RLHF Flow-GRPO implementation POC

[ifilipis]

Added reinforcement learning (Flow-GRPO) that seems to work quite universally across models.

It implements Flow-GRPO and lets you vote live and thus create a LoRA

Default parameters seem quite well tuned for quick results within a few iterations.
For extra fast results, set LR to 0.005 and behold the speed
The only difference in this implementation vs original: rewards are binary instead of relying on a ranking model

There's a new job type dropdown for creating Flow-GRPO tasks, and GRPO job has a voting interface that lets you generate samples and vote on them

Stuff yet to do:

• Manual checkpoints
• Reduce memory usage (Z-Image takes 40+ GB) and improve speed
• UI polishing
  • Prevent prompt settings from resetting on leaving the tab
  • Fix skip button color when selected
  • Add fixed/random seed
  • Improve progress indication when generating candidates
  • Maybe fix concurrency of training/generation
  • Improve candidate images layout
  • Add input images for editing models
• Keep testing the algorithm on all models