CUDA: flashattn head dim issues

[Green-Sky]

Removing this d_head check causes sd1 with fa to crash, since cuda does not support 40.

However, vulkan added support for this, so it works there just fine.
Incredibly, this makes the vulkan backend much faster for sd1+fa in comparison to cuda.

768x1024
cuda: 2.15s/it
vulkan: 1.20it/s 1.27it/s (with diffusion-conv-direct)

With it also having a better conv2d impl, it makes it now the obvious choice for sd1.
--diffusion-fa --vae-conv-direct --diffusion-conv-direct

Originally posted by @Green-Sky in cb1d975

(Commenting on a commit felt wrong.)

[Green-Sky]

Maybe we need to add an extra ggml_backend_supports_op() and see if we have to fall back...

[leejet]

I think #803 should fix the problem.

[Green-Sky]

@JohannesGaessler vulkan gained support for 40 and 160 with ggml-org/llama.cpp#15537 and it might be possible compiling the code paths for those k dims might just work without much code changes.

[JohannesGaessler]

It's not possible. As of right now the CUDA FA kernels have a minimum granularity of 16 for the head size, 64 for GPUs without tensor cores.

[Green-Sky]

I see, too bad.