feat(nvml-mock): nvidia-persistenced + Xid health-remediation loop

[giuliocalzo]

Priority: Low

Context. Failure injection (#328) emits Xid via the NVML event set, but the downstream remediation loop isn't exercised.

Gap. No nvidia-persistenced presence and no E2E covering NPD / device-plugin health then node cordon/drain on Xid.

Proposed scope.

• Optional fake nvidia-persistenced socket/presence.
• E2E: inject a critical Xid and assert the device-plugin health monitor (or NPD) marks the GPU Unhealthy / node gets cordoned.

Why. Validates the operational response to failures the mock can already inject. Low priority.