Add MXFP8 attention

[cyanguwa]

Description

1. Added MXFP8 support in FusedAttention backend (fwd+bwd)
2. Decoupled input/output format in APIs by introducing o_format, do_format, and dqkv_layout
3. Implemented mxfp8_quantize_fast_path() to quantize multiple tensors, pad/permute/swizzle the scale_invs for a more efficient quantization pipeline; added qkv_scale_inv_format, do_scale_inv_format to indicate scale_invs' format
4. Implemented multi_tensor_transpose_to_bhsd() to permute tensors from BSHD/SBHD to BHSD, with the TMA path optimized for FP16/BF16 dtype, D=192/128 cases, fallback_vec_aligned path for Byte, D=8/4, and a fallback_non_vec_aligned path for Byte, D=6
5. Implemented multi_tensor_pad_last_dim() to pad multiple tensors' D to %4 for rowwise and %128 for columnwise
6. Implemented multi_tensor_swizzle_row_scaling_narrow_k_kernel and multi_tensor_swizzle_col_scaling_narrow_m_kernel to optimize for small K/M dimensions
7. Added MXFP8PaddedSizes, pad_s_d_for_mxfp8, generateMatrixStridesWithFormat, generateMatrixStridesWithLayout for size/stride computation
8. Fixed bug in O/dQKV shape logic for MLA; added utility nvte_convert_qkv_shape
9. Refactored CP and added support for MXFP8 in cp_comm_type={'a2a', 'p2p', 'a2a+p2p', 'all_gather'}, with {'a2a', 'p2p', 'a2a+p2p', 'all_gather'} for MHA/GQA/MQA/MLA, {'a2a', 'all_gather'} for SWA, and {'a2a'} for sink attention
10. Fixed scale_inv_offsets calculation in GroupedTensor
11. MXFP8 attention requires cudnn-frontend v1.21+ and cuDNN 9.21+

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

Please see Description.

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

[greptile-apps]

Want your agent to iterate on Greptile's feedback? Try greploops.

[cyanguwa]

/te-ci L1

[cyanguwa]

/te-ci L1

[greptile-apps]

Greptile encountered an error while reviewing this PR. Please reach out to support@greptile.com for assistance.