Fix CUDA OOM during Krea2 block-by-block quantization

[thedivergentai]

Summary

• Fix CUDA OOM during Krea2 transformer quantization by synchronizing and flushing GPU memory after each block is offloaded to CPU.
• Ensure Krea2 starts quantization from a clean CPU baseline, including assistant LoRA weights when present.

Test plan

• Reproduced original failure path (OOM around block 25/28 during quantize_model)
• Verified all 28 blocks quantize successfully on a 24GB RTX 3090
• Confirmed Krea2 Raw training job completes model load and starts training