Use flash for dropout != 0

[mikekgfb]

PyTorch 2.0 (nightlies) now supports dropout > 0.0, so the check for dropout can be removed.
(And even if on back-level PyTorch builds, we can handle dropout > 0.0 where it will just fall back to sdpa_math() which is pretty much the same implementation based on the SDPA math that nanogpt uses today when flash is not available.)

Constraints to dispatch to Flash Attention are defined in use_flash_attention
https://github.com/pytorch/pytorch/blob/4973ca5e3e2c07311a879f49ac8983b7cae81a2d/aten/src/ATen/native/transformers/cuda/sdp_utils.h#L576
as follows:
https://github.com/pytorch/pytorch/blob/4973ca5e3e2c07311a879f49ac8983b7cae81a2d/aten/src/ATen/native/transformers/cuda/sdp_utils.h#L584-L592

[mikekgfb]

Fixed by #195