Model Modifiers - Perturbed Attention and Kohya Deep Shrink

[GitWilburn]

1. Added perturbed attention guidance Can improve attention at the cost of performance. Support for both base pass and hires pass (there is a breaking change for any existing app using run_model that SAG won't be included anymore, refer to CushyDiffusion for new method)
2. Added Kohya Deep Shrink Can adapt models to higher resolution than training, and helps with coherence for non-tiled hires fix. Example use: SD1.5 generate 2048x2048 using kohya scale of 4. Or upscale 512x512 4x on hires fix with less artifacts.
3. Added selection to SAG to choose if it applies to base or hires pass

[rvion]

Thanks 🙏 ! 🔥