Support hdimQK=192, hdimV=128 backward

[jacob-cursor]

Very happy to see 2a20412 which added support for the Deepseek attention configuration (qk_dim=192, v_dim=128) to FA3.

Currently, backward is not supported:

[rank1]:     dq, dk, dv, softmax_d, *rest = flash_attn_3_cuda.bwd(
[rank1]:                                    ^^^^^^^^^^^^^^^^^^^^^^
[rank1]: RuntimeError: out must have shape (batch_size, seqlen_q, num_heads, head_size)

The backward pass for this op would be really useful!

[tridao]

I'm personally not working on that soon, but I can review PRs if there's contribution