Forward serve logs + persist peak prefill/decode (monitor.sh)

launch.sh now starts a background monitor that tees the vLLM serve logs to
logs/serve.log (rotated 100MB) and tracks peak prefill/decode tok/s in
logs/peak.json, re-attaching across container restarts. Also adds docker
--log-opt rotation. Verified live: peak_decode 57.4 tok/s, peak_prefill 1401 tok/s.