Add ETA to every training log line + epoch summary

Fix: fast VAE encoding (bs=64), auto-limit large datasets, ~5x faster caching"

Add gradient checkpointing + auto batch size to prevent OOM on T4

Fix dataset: drop broken keremberke, use pure-parquet datasets only (cartoon default, Artificio/WikiArt for styles)"

Fix: add trust_remote_code=True for datasets with legacy loading scripts

Fix VAE: switch to madebyollin/sdxl-vae-fp16-fix (open, no auth needed, 4ch latent)

Add optimized training v2 with latent pre-caching for Colab

Fix: SDXL VAE (no login), streaming dataset, step-based training

Fix: streaming dataset (no full download), step-based training loop

Add training pipeline