scx_utils: use proper NVML API for GPU NUMA node lookup

[devnexen]

using the wrapper but still falling back to the slow(er) path in case. new attempt after #3108 only this time a new nvml-wrapper version had been released.

[devnexen]

cc @arighi

[htejun]

Please prove that this optimization is meaningful.

[devnexen]

To be honest, the win when I tried it was not that much performance wise mostly saving the FS operation time, I m digging the benchmark (hopefully still have it somewhere); but to summarize. The primary motivation is correctness rather than raw speed. The old code reverse-engineers a sysfs path from NVML's bus ID format by stripping a "0000" prefix — this is fragile and the FIXME comment in the
code acknowledges it may not always be 4 leading zeros.

numa_node_id() uses the proper NVML API (nvmlDeviceGetNumaNodeId) directly. The fallback to the sysfs path is kept for older NVML versions that don't support it.

[devnexen]

here it is, as gist. Locally it is only ~1.1 faster on average (I have only one nvidia gpu tough). Anyhow, using the now available direct wrapper NVML call alone, is a valid reason IMHO.