FLUX fine-tunes are now fast

This post details optimizations for running FLUX fine-tunes on Replicate, making them as fast as the base models. Key improvements include quantizing LoRAs to fp8, merging weights into the base model, and automatically increasing `lora_scale` by 1.5x when `go_fast=true` for better output quality. The post also introduces the ability to load LoRAs from Hugging Face, Civitai, and arbitrary HTTP URLs, and open-sources the improvements made to FLUX models.