fix: make qkv compatible with torch.compile in next diffusers release

[llcnt]

Description

The current main branch in diffusers is undergoing a large refactorization for attention computation.
The attention processors still exist but Flux and Wan have now a local version of their own processor. Also qkv fusing was changed: it is now taking place in the AttentionMixin class (not anymore in the transformer class itself).
Sister PR in pruna_pro is here.

Related Issue

The combination qkv_diffusers+torch_compile is taking forever to compute on the latest diffusers codebase

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• This change requires a documentation update

How Has This Been Tested?

Tested on Flux-dev with diffusers==0.34.0 and diffusers==0.35.0dev0
On Flux-dev, the generation time goes from 10.44s (original) to 5.12s (with qkv_diffusers+fp8+torch_compile).
The warm-up time (first inference) takes 56.16s.

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

Additional Notes

[johnrachwan123]

LGTM thanks a lot!

[nifleisch]

Looks good to me!