Separate Learning Rates for UNET and TE with Prodigy

Looking at Tensorboard graphs I noticed that Prodigy uses the same LR for both UNET and TE. Usually it's recommended to use lower LR for TE for better results (When I'm training LORAs with non-adaptive optimizers I usually use 1/10th or 1/5th of UNET's LR). In my personal experience of training LORAs for Illustrious-XL sometimes I've had results where LORA trained with Prodigy felt too well trained.
I looked at Prodigy's github repo and have found this Issue - [Lowering TE or Unet average only](https://github.com/konstmish/prodigy/issues/16), where @adefazio mentions that his [pull request](https://github.com/konstmish/prodigy/pull/9) for Prodigy - _"supports different LR values for different layers via setting the layer_scale for each in the param group:"_
Unfortunately, I don't have necessary knowledge to integrate these changes into sd-scripts to test if it improves training results.

Start by reviewing Prodigy's linked issue 16 and pull request 9 to understand the proposed per-parameter-group scaling. Then identify where sd-scripts configures Prodigy for UNET and TE training; done means those components can use separate learning rates and the training behavior can be evaluated against the reported concern.

Written by the indexing model from the issue text.