Add optimized training v2 with latent pre-caching for Colab

Fix: SDXL VAE (no login) + streaming dataset in notebook

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

Fix: use SDXL VAE (4ch, no login needed)

Fix: streaming dataset in notebook (no full download on Colab)

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

Add Colab training notebook

Add training pipeline

Add model architecture

Add README with architecture docs and usage guide

initial commit