fix: gracefully degrade util_pct when NVML unavailable, add nvidia-ml-py dep

feat: log GPU SM utilization (util_pct) alongside memory in [GPU_LOG]

feat: add GPU monitoring, named API endpoints, and stress-test script