PyTorch adds test for swizzled scales in scaled_mm compile

PyTorch has merged a test confirming that the Inductor compiler can now handle swizzled scales in the _scaled_mm_v2 operation. Previously, the lowering logic asserted that all swizzles must be NO_SWIZZLE, which broke torch.compile for scaled_mm operations using MX (microscaling) and NVFP4 recipes—these require NVIDIA's SWIZZLE_32_4_4 pattern on Blackwell B200 GPUs. The fallback fix landed in PR #186384; the new PR (#186948) adds the missing regression test to ensure the fix works.

The test compiles a blockwise MXFP8 scaled_mm with swizzled scales, compares the compiled result against eager execution, and asserts that the generated code calls the ATen kernel. It is B200-only (PLATFORM_SUPPORTS_MX_GEMM), matching the other blockwise tests. Authored with an AI assistant, this addition reinforces PyTorch's support for low-precision matrix multiplication on NVIDIA's latest architecture. For developers working with MXFP4/FP8 models, this means they can now fully leverage torch.compile to optimize performance without encountering compilation errors due to scale swizzle requirements.