Add torchvision.ops.deform_conv2d and opset19-specific tests

[f-fuchs]

Summary

Add mapping for torchvision.ops.deform_conv2d to ONNX DeformConv and add OpInfo-based test coverage.

Changes

• add torchvision.ops.deform_conv2d support in onnxscript/function_libs/torch_lib/ops/vision.py
• map PyTorch padding to ONNX pads
• infer and set group and offset_group
• support optional bias and mask

Tests

Added OpInfo coverage for:

• basic case
• bias
• mask
• nonzero padding
• grouped convolution
• multiple offset groups

Files:

• tests/function_libs/torch_lib/extra_opinfo.py
• tests/function_libs/torch_lib/ops_test_data.py

Opset handling

DeformConv requires ONNX opset 19+, while the existing torch-lib full-graph suite uses opset 18.

To keep the existing suite unchanged:

• keep the global torch-lib test baseline at opset 18
• add a dedicated deform-conv full-graph test path at opset 19
• exclude deform-conv from the generic opset-18 full-graph suite

Files:

• tests/function_libs/torch_lib/ops_test.py
• tests/function_libs/torch_lib/ops_test_common.py

Validation

python -m pytest tests/function_libs/torch_lib/ops_test.py -k deform_conv2d

Passed:

• 1 passed
• 1 skipped (expected for trace_only=True)

[f-fuchs]

@microsoft-github-policy-service agree