zlib: finish async convenience calls in one write

[anonrig]

The async convenience methods (zlib.deflate(), zlib.inflate(), zlib.gunzip(), ...) write the input to a Transform and then end it. That's two threadpool round trips per call, one for the data and one for the final flush, plus the stream machinery on top. In a CPU profile of benchmark/zlib/inflate.js method=inflate, the main thread sits idle about 71% of the time waiting on the threadpool.

We already have the whole input up front, so this hands it to the handle in a single write with the finish flush, the same way the *Sync methods do, and collects the output in the write callback. A call whose output fits in one buffer now takes one round trip. The work still runs on the threadpool.

These stay on the stream path:

• info: true, since the engine is handed back to the caller.
• A non-default flush, which can change the output bytes.
• zstdCompress(). A single ZSTD_e_end write makes frames 3–18 bytes larger than ZSTD_e_continue followed by ZSTD_e_end, so I left it alone.

The output doesn't change. For deflate, gzip, raw deflate and brotli, the old two-write sequence already produced the same bytes as the sync methods. I compared all the async methods plus gzipSync()/deflateSync() on four inputs (source text, text plus random bytes, a short string, empty) against a build of the base commit: 28/28 outputs are byte-identical, and each build decompresses the other's output.

There are two observable differences. The success callback now runs straight from the write completion, so async_hooks sees it in the ZLIB resource instead of a TickObject, which is how fs callbacks behave too. AsyncLocalStorage context is unchanged, and the new test checks it. Errors still arrive through the 'error' event like before. Also, this inside a non-arrow callback is still the internal engine, but it's never ended now, so writableEnded, writableFinished and readableEnded are false (it's still destroyed, with bytesWritten set), and info: true, the documented way to get the engine, keeps the old stream path.

Two new tests:

• test-zlib-convenience-async.js covers the contract: every input type, interop with the sync methods, invalid and truncated input, maxOutputLength, info, rejectGarbageAfterEnd, dictionaries, finishFlush, and that the callback never runs synchronously. It passes on the base commit too.
• test-zlib-convenience-async-round-trip.js uses async_hooks to check that a small call takes exactly one round trip. It fails on the base commit, which takes two.

Benchmarks

benchmark/compare.js, 40 runs, n=100000, against a build of the base commit with the same configure flags. The default method/inputLen are kept and only n is scaled down.

benchmark	run 1 (loaded machine)	run 2 (quiet machine)	A/A, quiet
inflate.js method='inflate'	+98.11% ***	+94.91% *** ±0.59%	+0.10%
deflate.js method='deflate'	+62.96% ***	+66.27% *** ±2.34%	+0.34%

--max-regression 5 over deflate.js, inflate.js, pipe.js and creation.js (20 runs) passes. The sync methods, createDeflate writes, pipes and stream creation all stay within about 2%.

compare.js, run 2 (40 runs)

                                                         confidence   improvement   accuracy (*)    (**)   (***)
zlib/inflate.js n=100000 inputLen=1024 method='inflate'  ***            +94.91 %   ±0.59%  ±0.78%  ±1.02%

                                                         confidence   improvement   accuracy (*)    (**)   (***)
zlib/deflate.js n=100000 inputLen=1024 method='deflate'  ***            +66.27 %   ±2.34%  ±3.11%  ±4.06%

compare.js --max-regression 5 (20 runs)

                                                               confidence   improvement   accuracy (*)    (**)   (***)
zlib/deflate.js n=100000 inputLen=1024 method='createDeflate'                  +0.84 %   ±2.03%  ±2.73%  ±3.58%
zlib/deflate.js n=100000 inputLen=1024 method='deflate'        ***            +75.51 %   ±1.87%  ±2.51%  ±3.31%
zlib/deflate.js n=100000 inputLen=1024 method='deflateSync'                    +0.12 %   ±1.06%  ±1.42%  ±1.86%

                                                             confidence   improvement   accuracy (*)    (**)   (***)
zlib/inflate.js n=100000 inputLen=1024 method='inflate'      ***            +95.85 %   ±3.53%  ±4.79%  ±6.43%
zlib/inflate.js n=100000 inputLen=1024 method='inflateSync'                  -1.56 %   ±3.28%  ±4.41%  ±5.84%

                                                                        confidence   improvement   accuracy (*)    (**)   (***)
zlib/pipe.js algorithm='gzip' type='string' duration=1 inputLen=1024                    +1.74 %   ±2.32%  ±3.13%  ±4.18%
zlib/pipe.js algorithm='brotli' type='string' duration=1 inputLen=1024                  +0.12 %   ±3.43%  ±4.59%  ±6.04%
zlib/pipe.js algorithm='zstd' type='string' duration=1 inputLen=1024                    -0.15 %   ±1.40%  ±1.88%  ±2.47%
zlib/pipe.js algorithm='gzip' type='buffer' duration=1 inputLen=1024                    -0.90 %   ±1.75%  ±2.35%  ±3.10%
zlib/pipe.js algorithm='brotli' type='buffer' duration=1 inputLen=1024                  +0.74 %   ±1.50%  ±2.06%  ±2.79%
zlib/pipe.js algorithm='zstd' type='buffer' duration=1 inputLen=1024      *             -1.54 %   ±1.16%  ±1.55%  ±2.05%

                                                                   confidence   improvement   accuracy (*)    (**)   (***)
zlib/creation.js n=500000 options='true' type='Deflate'                            +0.49 %   ±1.14%  ±1.53%  ±2.01%
zlib/creation.js n=500000 options='false' type='Deflate'                           -0.43 %   ±0.72%  ±0.96%  ±1.26%
zlib/creation.js n=500000 options='true' type='DeflateRaw'           *             -2.01 %   ±1.55%  ±2.10%  ±2.81%
zlib/creation.js n=500000 options='false' type='DeflateRaw'                        +0.81 %   ±1.64%  ±2.23%  ±2.98%
zlib/creation.js n=500000 options='true' type='Inflate'                            -0.19 %   ±1.01%  ±1.36%  ±1.80%
zlib/creation.js n=500000 options='false' type='Inflate'                           +1.17 %   ±1.68%  ±2.27%  ±3.02%
zlib/creation.js n=500000 options='true' type='InflateRaw'                         -1.01 %   ±2.40%  ±3.23%  ±4.25%
zlib/creation.js n=500000 options='false' type='InflateRaw'                        +0.72 %   ±1.35%  ±1.82%  ±2.41%
zlib/creation.js n=500000 options='true' type='Gzip'                               +0.91 %   ±1.44%  ±1.93%  ±2.56%
zlib/creation.js n=500000 options='false' type='Gzip'                              +0.34 %   ±1.60%  ±2.16%  ±2.87%
zlib/creation.js n=500000 options='true' type='Gunzip'                             -0.17 %   ±1.15%  ±1.55%  ±2.04%
zlib/creation.js n=500000 options='false' type='Gunzip'                            +0.61 %   ±1.74%  ±2.33%  ±3.06%
zlib/creation.js n=500000 options='true' type='Unzip'                              -1.05 %   ±1.72%  ±2.31%  ±3.06%
zlib/creation.js n=500000 options='false' type='Unzip'                             +0.09 %   ±0.97%  ±1.31%  ±1.74%
zlib/creation.js n=500000 options='true' type='BrotliCompress'                     -2.17 %   ±3.48%  ±4.74%  ±6.39%
zlib/creation.js n=500000 options='false' type='BrotliCompress'                    -1.24 %   ±2.90%  ±3.92%  ±5.25%
zlib/creation.js n=500000 options='true' type='BrotliDecompress'                   -0.96 %   ±2.39%  ±3.21%  ±4.23%
zlib/creation.js n=500000 options='false' type='BrotliDecompress'                  -1.13 %   ±2.70%  ±3.62%  ±4.76%
zlib/creation.js n=500000 options='true' type='ZstdCompress'                       +0.13 %   ±1.42%  ±1.91%  ±2.52%
zlib/creation.js n=500000 options='false' type='ZstdCompress'                      -0.24 %   ±1.38%  ±1.85%  ±2.44%
zlib/creation.js n=500000 options='true' type='ZstdDecompress'                     -0.28 %   ±2.36%  ±3.17%  ±4.21%
zlib/creation.js n=500000 options='false' type='ZstdDecompress'                    +0.13 %   ±1.49%  ±2.00%  ±2.63%

Below, Chrome talks to two local servers, one on the base build and one on this PR. It decodes zlib.gzip()/zlib.deflate()/zlib.brotliCompress() responses, and sends CompressionStream uploads that the servers decode with zlib.gunzip()/zlib.inflate(). The payloads match on both, the round trip count drops from 2 to 1, and sequential zlib.inflate() calls get faster.