[Hopper] Support sink attention (learnable sink)

[aoxy]

Description

This PR adds support for sink attention (learnable sink) in FlashAttention-3 (Hopper architecture), matching the functionality already available in FlashAttention-2.

Sink attention allows a learnable parameter per head to be added to the attention scores before the softmax operation. This provides a "sink" that can absorb attention probability, which has been shown to be useful for models like GPT-OSS.

Key Changes:

• Core Implementation: Added support for learnable_sink in the Hopper forward and backward kernels.
• Python Interface: Updated flash_attn_func, flash_attn_varlen_func, and flash_attn_with_kvcache to accept the learnable_sink tensor.
• Backward Pass: Implemented the gradient calculation for learnable_sink, enabling end-to-end training of the sink parameters.
• C++/CUDA Bindings: Updated the mha_fwd and mha_bwd bindings in hopper/flash_api.cpp to handle the new learnable_sink and dsink tensors.
• Compatibility: The implementation is consistent with the sink attention support in FAV2.

Testing:

• Added comprehensive tests for sink attention in hopper/test_flash_attn.py.
• Updated the reference attention implementation in hopper/test_util.py to include sink attention logic.
• Verified both forward and backward passes across various configurations (MHA, MQA, GQA, causal, varlen, KV cache, etc.).
• All tests for sink attention have passed successfully.

I have already added the tests to tests/ and confirmed they pass.