Support Context Parallel for Multi Latent Attention (MLA)

[yuzhongw-nvidia]

Description

#1561 has already fixed the issue that the function AttnFuncWithCPAndKVP2P does not support MLA (Multi-latent attention). Specifically, #1561 pad tensor v to head_dim_qk and convert MLA to normal attention. This PR improves #1561 by removing the padding and using MLA kernels to reduce the communication and computation overhead.

Many thanks to SuperCB from xiaohongshu and RandMist from wechat team for their contributions.

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

Please list the changes introduced in this PR:

• Change A
• Change B

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