usage of xformers in gen_img_diffusers.py

Hi, I appreciate the clever structure of the inference script!

I notice that you only change the forward method in CrossAttention but not in other AttentionBlock when uising the xformers, which is difference from the diffusers. [code](https://github.com/kohya-ss/sd-scripts/blob/main/gen_img_diffusers.py#L384). And it seems works better than diffusers too.

Why it works better? Thanks a lot!

Read gen_img_diffusers.py around line 384, then compare the CrossAttention forward path with the other AttentionBlock usage mentioned in the issue and the linked diffusers implementation. Reproduce or inspect the xformers paths to identify the behavioral difference; done means documenting a supported explanation of why the results differ, if one can be established.

Written by the indexing model from the issue text.