Fix NVMLError_Unknown: init NVML once, auto reset+retry on failure

[tugrultamerc]

The app called nvmlInit/nvmlShutdown on every request from multiple threadpool workers. After a host driver reload/update this left NVML in a bad state and /gpu returned 500 with NVMLError_Unknown.

• Initialize NVML once per process (thread-safe), shutdown at exit
• On NVMLError during gather: shutdown+re-init and retry once
• On init failure: attempt one reset before giving up
• Return schema-valid status=error payload instead of an unhandled 500
• Add smoke tests covering outage, recovery, and retry behavior