Add ARB for training

[Yidhar]

In the training environment, there is no automatic clustering according to the resolution ratio, that is, buckets. I think this feature is needed. Training at the same resolution consumes much more memory per batch size than sdxl. The unet+ prior model of Kandinsky2.2 should total 2.2b slightly less than sdxl. Maybe there are some performance issues? The second is when using the unload model feature in the Settings. lora failed to be properly released from video memory (and possibly other model caches). When I had to uninstall the model repeatedly and then load the model with lora to complete the sampling. More and more content stays in video memory. Until I had to close the app completely to erase the memory

[seruva19]

Aspect radio bucketing was already in my personal feature wishlist, so sooner or later I will implement it. Cannot provide exact timeline, though, since I don't have much time to work on this project currently.
As for training, I simply used training scripts from the diffusers' repo, so there surely must be some room for optimization.
And regarding the issues with LoRA and model unloading, I will try to fix them.

Thank you for your feedback; I appreciate the time you spent trying out this app.

[seruva19]

Probably will be implemented as part of the implicitly planned LoRA training for KD 4.0, so still not closing it.