[torchlib] Fix torchvision::roi_align lowering to accept 7-arg schema

[FraGirla]

Dynamo ONNX export calls torchvision::roi_align with (input, rois, spatial_scale,
pooled_h, pooled_w, sampling_ratio, aligned), but torchlib’s lowering expected
output_size and rejected the extra args.

Update torchvision_roi_align to take pooled_h/pooled_w directly and add an OpInfo
input_wrangler to adapt wrapper-style roi_align(input, boxes, output_size, ...).

Refs pytorch/pytorch#175732
Test: pytest tests/function_libs/torch_lib/ops_test.py -k roi_align -v

[FraGirla]

@microsoft-github-policy-service agree

[justinchuby]

Please update the extra_opinfo file instead of creating a wrangler.

[FraGirla]

Done in the last commits

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 71.77%. Comparing base (ef2bc22) to head (d1cb271).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2830      +/-   ##
==========================================
- Coverage   71.78%   71.77%   -0.01%     
==========================================
  Files         239      239              
  Lines       28989    28988       -1     
  Branches     2859     2859              
==========================================
- Hits        20809    20806       -3     
- Misses       7209     7210       +1     
- Partials      971      972       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[justinchuby]

Still want to check when this was changed. Since it is a breaking change we need to define the function conditionally based on pytorch version.

[FraGirla]

I checked versions: this fix targets the torch.export/Dynamo ONNX exporter path (torch.onnx.export(..., dynamo=True)), which is available in recent PyTorch releases (docs mention >= 2.5). The change simply aligns the torchlib lowering with the underlying operator schema for torchvision::roi_align (7 args: input, rois, spatial_scale, pooled_height, pooled_width, sampling_ratio, aligned).

The operator itself has been 7-arg for a long time; what’s new is that the Dynamo ONNX exporter path now hits the operator-level call directly, while the previous torchlib lowering matched the Python wrapper signature (output_size tuple), so the mismatch only surfaced once this path got exercised more.

Also fixed lintrunner issues in a follow-up commit.