checkpoint_manager.build_model: load tokenizer bundled with the checkpoint (fixes FileNotFoundError when NANOCHAT_BASE_DIR/tokenizer isn't pre-populated, e.g. standalone deployments loading straight from the HF snapshot dir)

Add nanochat/engine.py for KV-cache batched inference

Point inference code at final checkpoint (step 1291)

Add self-contained inference code (inference.py, nanochat/, tokenizer/, config, README) for step 1000

Upload folder using huggingface_hub

initial commit