[PyTorch] Enable e5m3 fused GEMM kernels from cuDNN

[kainzhong]

Description

Route NVFP4 with e5m3 SF GEMM to cuDNN. Only applied to basic linear and fused grouped linear for now

TODO:

• grouped_gemm_quant_wrapper_sm100 doesn't support accumulation, so I have to turn off with cfg.model.gradient_accumulation_fusion.
• I have to pad MLA's kv_a_proj_with_mqa due to cuDNN-FE's grouped_gemm_quant_wrapper_sm100 requiring M to be divisible by 256. Maybe I can also try to use grouped_gemm_wgrad_wrapper_sm100 instead which doesn't require divisibility

The numeric seems wrong when I train DeepSeek V3. I'm still debugging it

Fixes # (issue)

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

• general_gemm now goes to cuDNN-FE's grouped_gemm_quant_wrapper_sm100 when NVFP4 with e5m3 SF
• grouped_mlp.py uses cuDNN-FE's Rubin kernels when NVFP4 with e5m3 SF
• Pass scale's dtype to C++ from _cast_master_weights_to_nvfp4_2d otherwise quantization will default to e4m3 and emits wrong result

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes