Add FLUX finetuning

[angeloskath]

This adds a dreambooth example for FLUX-dev finetuning. It still needs a README and possibly some more cleaning but it works well for the dev model. The following is a list of things to do in the future

• Prior preservation
• Ensure it works for schnell (maybe find hyper parameters)
• QLoRA for machines with smaller memory
• Check if we can avoid T5 padding for some extra speed

Distributed training scales almost perfectly, most of my training was with 8 nodes.

[akkikumar72]

Awesome work 🎊 Will be nice to have a readme for this, so we can try/test it in weekend

[angeloskath]

@akkikumar72 I totally agree. I am AFK for a few hours but I will write the README and get it merged tonight (PST).

[nattaylor]

This is amazing. Any idea when you might have QLoRA implemented? :)