[WIP] Experimental FlyDSL GEMM backend for TE PyTorch (BF16/FP16/FP32/FP8/MXFP8)

[aris134]

Summary

This PR introduces an experimental FlyDSL GEMM backend for the TransformerEngine PyTorch frontend, following PR #667

The initial implementation wires FlyDSL into the existing GEMM dispatch path and establishes the infrastructure needed to invoke FlyDSL kernels from TransformerEngine. FlyDSL is integrated as an optional dependency and is installed when the corresponding build option is enabled during TransformerEngine installation. At runtime, the backend is opt-in and selected via the NVTE_USE_FLYDSL=1 environment variable; otherwise, the existing TransformerEngine GEMM backend is used unchanged.

The FP16/BF16/FP8/MXFP8 GEMM paths have been integrated and verified against the existing TE implementation using a standalone end-to-end validation.

Remaining items

• Extend MXFP8 support beyond the initial implementation to the remaining TE GEMM configurations.
• Add backend selection tests and numerical validation covering all supported datatypes.
• Add performance benchmarking and characterization against the existing TE backends (hipblas, triton).