Metal backend: store weights outside .so

[manuelcandales]

This pull request refactors how model weight constants are packaged and loaded in the Metal backend. The main change is separating weight constants from the shared object (.so) file, storing them as a binary blob on disk, and updating the runtime to load these weights from the blob. This speeds up loading time.

[manuelcandales]

Stack from ghstack (oldest at bottom):

• Metal backend: Add MPSGraph caching #15346
• Metal backend: eliminate memory leak #15343
• Metal backend: track tensors #15342
• -> Metal backend: store weights outside .so #15341

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/15341

• 📄 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, 3 Pending, 2 Unrelated Failures

As of commit d8e6d13 with merge base 5d71c9b ():

NEW FAILURES - The following jobs have failed:

• pull / unittest-nxp-neutron / linux-job (gh)
  RuntimeError: Command docker exec -t 22faf4bcc032567dad5b632f7ccda4d44719541f890db61c64030c9a3a627912 /exec failed with exit code 1
• Test CUDA Builds / benchmark-voxtral-cuda / linux-job (gh)
  RuntimeError: Command docker exec -t 1e1ed95bed54d090de8e292fb2234512e70fdde08750457f05e2bada5a65a3ff /exec failed with exit code 2
• Test CUDA Builds / export-gemma3-cuda-non-quantized / linux-job (gh)
  AttributeError: 'NoneType' object has no attribute 'reorder_iter_loops'

BROKEN TRUNK - The following jobs failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

• pull / unittest / linux / linux-job (gh) (trunk failure)
• pull / unittest-editable / linux / linux-job (gh) (trunk failure)

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

[larryliu0820]

Is there a reason why we need metal_backend.py? It feels very much like cuda_backend.py

[manuelcandales]

Yes, I agree. Next week we should begin refactoring, and moving most of this code to a new aoti_backend.py. Same with most of the code in metal_backend.cpp/cuda_backend.cpp