perf(cpp): share a writer-owned schema cache across tree and table writes

[ColinLeeo]

Repeated writes currently resolve each measurement name through a per-device schema map. This PR reuses those lookups through one bounded cache owned by TsFileWriter, covering tree records, tree tablets, and table FIELD columns.

Related to #885 and the earlier tree-only implementation in #934. This is an alternative implementation based on the current develop branch; the initial cache regression cases build on #934.

Changes

• Keep an LRU of at most 64 device entries per writer, with a fast path for the most recent device. Entries hold non-owning schema pointers and preserve the input column order; positional names are checked on every call, and unresolved columns are looked up again after registration.
• Share the cache across aligned/non-aligned tree writes and table FIELD lookup. Clear it before schema destruction; flush preserves the schema objects. Eviction does not invalidate the chunk writers already saved in a table write context.
• Keep omitted aligned fields synchronized with time rows, pages, and chunks. Cache the omitted-field list so repeated full-schema writes avoid rescanning all registered fields. Fields registered before the first flush receive the preceding empty pages and null positions; adding a new aligned field after the device's first flush returns E_NOT_SUPPORT rather than producing incomplete historical alignment.
• Read the existing one-varint empty aligned-page representation and reset the corresponding reader state. Preserve a single empty aligned value page during crash recovery so later valid chunks are not truncated.

The entry-count bound is 64 devices, not a byte limit: each entry retains vectors sized for its device's schema. This does not add support for concurrent public calls on one writer; internal parallel encoding continues to use the existing thread pool.

Validation

• macOS arm64, Apple Clang 21; Release build with bundled compression dependencies and LZMA2 enabled.
• Full C++ test binary: 957 passed, 3 skipped because their external fixture environment variables were unset; 10 existing disabled tests were not run.
• Explicitly instrumented Debug ASan/UBSan run: all 92 targeted tests completed successfully with no ASan findings (detect_leaks=0). UBSan reported existing null-pointer arithmetic at common/allocator/page_arena.h:92; a standalone PageArena::alloc(2048) followed by alloc(8) with a 1024-byte page reproduces it without any writer or cache. The allocator is unchanged from the base, and this is not a clean UBSan run.
• All ten changed C++ files pass clang-format 17.0.6 verification and git diff --check.

The regression suite checks row values and timestamps across repeated writes, changed names/order/width, late registration, multiple devices and writers, destroy/re-init, eviction beyond 64 devices, table TAG/FIELD layouts, sparse aligned fields, serial/parallel page boundaries, and recovery preserving an empty-page backfill and the following valid chunks.