Feat textual inversion plugin v2

[damian0815]

With this plugin you can target specific token embeddings in the text encoder, and/or add new ones. The embeddings get written into the model's text encoder and saved out - not the most efficient way to share them, but useful if you want to laser-focus training of specific or custom tokens in your model.

Documentation (such as it is) is in plugins/textual_inversion.py. TL;DR is - first edit plugins/textual_inversion.json to set up your tokens and initialization states. enable text encoder training and disable unet training. then edit optimizer.json to freeze all TE layers and TE final layer norm. the plugin checks for sane config when it's used.

you'll want to use a very high LR for the TE. my last test worked best at 2e-2.