FLUX fine-tunes are now fast

What this post added

This post details optimizations for FLUX fine-tunes, making them as fast as base models. This is achieved by quantizing LoRAs to fp8, merging weights into the base model, and automatically increasing `lora_scale` by 1.5x when `go_fast=true`. Additionally, support has been added to load LoRAs from Hugging Face, Civitai, and arbitrary HTTP URLs. The underlying optimizations for the base models utilize `torch.compile` and fast CuDNN attention kernels.