Add mm operator support

[voltjia]

Summary

This PR implements aten::mm for two-dimensional, same-device float32 Infini tensors by calling InfiniOps Gemm implementation index 0 on the current stream.

The initial path supports contiguous inputs, transposed non-overlapping dense weights, empty dimensions, and input/output storage lifetime tracking. It intentionally rejects unsupported dtypes, devices, ranks, and overlapping layouts instead of falling back through CPU.

This enables torch.nn.Linear(..., bias=False) inference through PyTorch's existing transpose-and-mm path. Bias, mm.out, training, backward, and propagation of torch.set_float32_matmul_precision("highest") remain outside this PR.

This PR is stacked on #37, which is stacked on #36. The intended merge order is #36, #37, and then this PR.

Testing

• Ruff 0.15.21 and clang-format 19.1.4 pass.
• An installed wheel built against PyTorch 2.12.0+cpu passes the focused suite with 11 passed, 1 skipped, and 1 xfailed, and the full suite with 181 passed, 43 skipped, and 3 xfailed.
• An installed wheel built against PyTorch 2.13.0+cpu passes the focused suite with 11 passed, 1 skipped, and 1 xfailed, and the full suite with 181 passed, 43 skipped, and 3 xfailed.
• An installed wheel built against PyTorch 2.13.0+cu130 passes the focused suite on an NVIDIA A100-SXM4-80GB with 11 passed, 1 skipped, and 1 xfailed.
• The NVIDIA validation uses driver 580.105.08, CUDA toolkit 13.1.80, CUDA architecture 80, InfiniRT commit 95c70080f9551e61241110497d163dfcdf9dc7e7, and InfiniOps commit 296271487beb594a248fd463e5fff14f7ab74293.
• The Infini results for contiguous and transposed-weight inputs match official torch.cuda.mm under PyTorch 2.13's default high float32 matrix-multiplication precision.
• The acceptance suite runs an actual bias-free torch.nn.Linear module after moving its weight and input to the infini device.