[executorch][native] Map .ptn packages instead of copying them

[SS-JIA]

Stack from ghstack (oldest at bottom):

• -> [executorch][native] Map .ptn packages instead of copying them #22532
• [executorch][native] Add the .ptn package facade and its inspector #22471
• [executorch][native] Add a safetensors index reader #22470
• [executorch][native] Add a zip reader for .ptn packages #22469
• [executorch][native] Add a minimal JSON reader for package metadata #22468
• [executorch][native] Add deserializer bridge (Program -> Method) #22467
• [executorch][native] Claim getitem in the partitioner #22466

Package read the whole .ptn into a std::vector<uint8_t>, so every constant
was heap-resident for the package's lifetime whether or not anything touched it.
Fine for a 14 MB mobilenet_v2; disqualifying for a model whose weights are the
reason constants ship out-of-line in the first place.

Adds OwnedBytes: owning, read-only bytes that are either a heap buffer or a
read-only file mapping, behind one span(). Package holds one of these instead
of a vector, so the backing becomes a load-time argument rather than a property
of the type. Package::load(path) maps by default; pass use_mmap=false to read
into the heap. load_file and map_file collapse into that one entry point, and
the read_file that was duplicated in Package.cpp and ptn_inspector.cpp
collapses into OwnedBytes::from_file.

The mapping covers the whole file from offset 0, which is deliberately simpler
than ExecuTorch's MmapDataLoader: mapping a sub-range needs page-rounding on the
way in and again in the unmap callback, since only the interior pointer survives
in the buffer, whereas a whole-file map is already page-aligned. That also means
alignment is unchanged from the heap path -- a page-aligned base is more aligned
than new[]'s, so every span keeps the alignment its file offset implies. The
caveat in SafeTensorsReader.h about unpadded safetensors payloads applies
equally to both modes; removing it is an AOT change.

MAP_SHARED, so two processes running the same model share physical pages, the
way MmapDataLoader does. A zero-length file yields empty heap bytes, since
mmap rejects a zero length. Platforms without mmap throw rather than failing
to compile.

ptn_inspector maps by default and takes --nommap to opt out. It cannot use
Package::load(path), because it also accepts a bare .ptg and so must inspect
the leading bytes before it knows whether to build a package at all;
re-acquiring the file inside the path overload would read it twice under
--nommap. That overload therefore has no caller yet -- it is the entry point
the runtime will use -- though both halves it forwards to are exercised here.

One behaviour difference worth knowing: a mapping sees later edits to the file on
disk, a heap read does not.

Differential Revision: D118480655

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22532

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures

As of commit 6073d95 with merge base 2c1da32 ():

NEW FAILURES - The following jobs have failed:

• pull / android / build-android (gh)
• pull / test-arduino-library / run / arduino-library (gh)
  RuntimeError: Command docker exec -t 0f477d0c361b202a3b8f828b490709e6400bd7636615f5b741b4957160ab8e46 /exec failed with exit code 1
• pull / test-samsung-models-linux / linux-job (gh)
  AttributeError: type object 'Test_Milestone_MobileBertFinetune' has no attribute 'model_cache_dir'

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.

[digantdesai]

Review automatically exported from Phabricator review in Meta.