Arm backend: Fuse canonical flow-offset grid sampling

[YufengShi-dudu]

Recognize grid_sample calls that add a two-channel flow slice to an align-corners base grid:

base_grid ---------\
                     add -> permute -> grid_sample(image, grid)
flow[:, i:i+2] ----/

Replace the complete expression with:

image, flow -> fused flow-offset custom shader

This avoids materializing the normalized sampling grid at runtime.

Require SNORM-compatible image and output quantization ranges. Vulkan's signed normalized 8-bit image format maps both -128 and -127 to -1.0, so only fuse tensors using the range [-127, 127]. The flow tensor keeps the full int8 range because the shader reads it directly as a tensor.

If glslc is unavailable or compilation fails, leave the graph unchanged for generic grid-sampler lowering.

Cover supported rewrites, invalid patterns, SNORM ranges, and compiler fallback behavior.

Authored with Codex.

Change-Id: Ia19b4387c169bd24f8db837a5bfee7c91d4e55ce

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21775

• 📄 Preview Python docs built from this PR

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

❌ 8 Pending, 1 Unclassified Failure

As of commit b34ec8b with merge base 8de0262 ():

UNCLASSIFIED FAILURE - DrCI could not classify the following job because the workflow did not run on the merge base. The failure may be pre-existing on trunk or introduced by this PR:

• trunk / test-huggingface-transformers-macos (olmo-1b|coreml_fp32_gpu|--quantize) / macos-job (gh) (this job did not run on the merge base, so DrCI cannot tell whether the failure is pre-existing)
  pytorch_tokenizers

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

[robell]

Extensive internal review.

[YufengShi-dudu]

unrelated failures