[PyTorch] Support FA3 MLA CP feature

[zhujian19891203]

Description

Because flash-attention #1604 has already support hdimQK != hdimV backward, so we can support FA3(Flash Attention 3) backend for MLA (Multi-latent attention). #1604 allows us to skip explicit padding & unpadding hdimV to use FA3 as attention backend, and it can bring performance benefits.

Test Results:

1. I add some unit tests for FA3 MLA CP, test_fused_attn.py and test_fused_attn_with_cp.py passed.
2. I use the Megatron-LM framework to run the 16B DeepSeek model, where TP=CP=2, hdimQK=192, hdimV=128, FA3's MFU is basically the same as cuDNN attention, and the overall loss curves of the two are exactly the same.

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

1. Update FA3 to commit-id 3ba6f82 (tag 2.8.0.post2 with compile error fixed)
2. Update get_attention_backend method because FA3 support MLA now
3. Support FA3 MLA for CP module
4. Add unit tests for FA3 MLA CP
5. Update attention doc

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