Rewrite Dockerfile: runpod/base + uv sync from lockfile

Persist HF_TOKEN to cache file in entrypoint (survives /start.sh env changes)

Runner entrypoint: delegate SSH to Runpod's /start.sh, fix monitor script

Fix SSH: generate host keys, use 'ip' field from runpodctl

Start sshd in runner entrypoint for monitoring without losing auto-stop

Support PAWN_CMD env var in runner entrypoint for runpodctl

Safetensors migration, checkpoint integrity, and multi-model training. (#1)