v3: patch_size=4 (64 tokens), 2 core layers, iters [2,3,4], ~16min total training

Fix: force_download=True to bypass HF cache, add assertion checks

v2: Use pre-trained SD-VAE, fix all bugs, pre-cache everything, massive speedup

Perf fix: cache Manhattan dist, optimize scan, pre-cache CLIP, fix deprecated AMP API

Add complete Colab/Kaggle training notebook