[torchlib] Add aten::_grouped_mm ONNX lowering

[TANGBUDU]

Summary

Adds a torchlib lowering for aten::_grouped_mm, addressing #2795.

• lowers 3D x 3D grouped matmul directly to ONNX MatMul
• supports the offset-based 2D x 3D, 3D x 2D, and 2D x 2D layouts by unrolling groups when the group count is statically known
• preserves dynamic shapes for the regular 3D x 3D case
• keeps the existing PyTorch constraints around offsets, bias, and output dtype
• guards the new OpInfo/e2e tests so stable PyTorch releases that do not yet expose _grouped_mm remain unaffected

For an offset tensor whose number of groups is itself dynamic, this PR intentionally does not build a generic loop. A future follow-up could lower that case with Scan/Loop; the static-group unrolling here follows the direction discussed in #2795.

Tests

• pytest tests/function_libs/torch_lib/ops_test.py -k grouped_mm -q
  • 4 offset/layout subtests pass
• pytest tests/function_libs/torch_lib/e2e_ops_tests.py -k grouped_mm -q
  • 2 passed
• lintrunner -a
  • no lint issues
• git diff --check
  • clean
• stable PyTorch 2.7.1 compatibility check
  • _grouped_mm tests are correctly excluded/skipped

I also ran the full torch-lib e2e file on the current PyTorch nightly. It has 5 unrelated failures (STFT graph API, dynamic unbind shape constraints, and SDPA numerical mismatch); all 5 reproduce unchanged on a clean upstream/main checkout with this patch stashed.

[TANGBUDU]

@microsoft-github-policy-service agree