Cudnn frontend sdpa wrong assertion in attention module with d_bias

[Skylion007]

Describe the bug
A clear and concise description of what the bug is.

Cudnn_frontend asserts on d_bias, even when the d_bias isn't used. See this PyTorch issue: pytorch/pytorch#137347 When upgrading the CUDNN_Frontend, code that used to work fine no longer works, and the assertion does not seem appropiate:

RuntimeError: cuDNN Frontend error: For cuDNN version below 9.5.0, dBias not support s_q/s_kv which aren't multiple of 64

pytorch/pytorch#137347

Expected behavior
A clear and concise description of what you expected to happen.
Previously working code would remain working. The assertion appears to be incorrect as the dbias isn't used:

System Environment (please complete the following information):

• cudnn_frontend 1.7.0

To Reproduce
Steps to reproduce the behavior:

1. Update PyTorch and update the submodule to 1.7.0
2. Run the example in cuDNN dBias error starting on 10/2 nightly pytorch/pytorch#137347
3. Watch the cudnn error propogate

Additional context
Add any other context about the problem here.

[Skylion007]

Ping @eqy since you were working on a fix

[eqy]

IIUC cuDNN is targeting a fix in a few days-week in an upcoming release