Validate on an NVIDIA host: every metric, Xids, NVLink, MIG

[moezdil]

The NVIDIA provider is tested against a fake NVML library in CI and has not run on real hardware. Run accel --once --json --vendors nvidia and the interface on a host with NVML, and check every metric, Xid events, NVLink counters, and MIG slices against nvidia-smi. A report through the hardware issue template is enough to start.

[moezdil]

Progress: ran accel on a real NVIDIA H100 SXM (Nebius, driver 570.211.01). Every metric checked against nvidia-smi matched, idle and under a real 2 GiB CUDA allocation (device-wide and per-process). Found and fixed a real bug along the way (#34): the plain nvmlDeviceGetMemoryInfo call folded reserved memory into Used on this driver; accel now prefers the v2 call.

Still open on this issue: NVLink (needs 2+ GPUs) and MIG (needs a MIG-enabled instance); this host had neither. Also could not reach an L40S (Nebius had no capacity for it at the time) for a second GPU generation.