Update README with v3 architecture info and speed comparison

Add v3 training script (self-contained, works with HF Jobs)

Add BokehFlow v3: 4400x faster (conv recurrence replaces sequential loop)

v4: Fix 429 — use snapshot_download with exact allow_patterns, no raw HTTP. 500 scenes in 80s, zero errors.

v3: Fix 429 rate limit — download-then-train with retry/backoff, disk-based DataLoader, 8 threads, HF token auth

v2: Zero-download streaming notebook — 3s startup, 0 disk, images fetched on-demand via HTTP

Add Colab/Kaggle training notebook — just run all cells

Add code audit — 10 issues found with fixes

Add detailed architecture design document

Add BokehFlow implementation - complete PyTorch architecture

Add comprehensive README with architecture details, research survey, and documentation

initial commit