[Relax] Integrate cuDNN attention

[vinx13]

This integrates cuDNN attention kernels to BYOC.
A dependency of cudnn_frontend is added.
The cuDNN attention kernel supports fused qkv in BS3NH and SBN3H layouts.

cc @sunggg @masahi @yongwww @tqchen

[sunggg]

Thank you @vinx13 for the new addition!
Overall, looks good to me.
Would you describe high-level strategy for attention somewhere? (e.g., when to offload cudnn, cutlass, TIR, etc.)
If this is about landing machinery rather than such offloading decision, would appreciate if you can provide some recommendations.

[vinx13]

The new attention can be applied via cudnn BYOC. The decision of which BYOC backend (cudnn, cutlass) to use is left to the users. cudnn is likely to perform better on H100 as it has specific optimizations

[masahi]

I remember cuDNN attention supports fp8, would be interesting to support that too.