Add Deep-NanoGPT experiment (Phase 1 & 2): resumable training, inference, 72-layer models

671ce97