Expand `cuda.core.system` to cover NVML functionalities

[leofang]

Tentative design considerations/constraints:

1. Must not run cuInit(0) or initialize CUDA if all users need is querying via NVML
2. Cover nvmlDeviceGetCpuAffinity, nvmlDeviceGetMemoryInfo, nvmlDeviceGetCount, ...

[leofang]

Blocked by #863 (unless we're willing to use nvidia-ml-py as stop-gap).

[mdboom]

I would say all of the "core" functionality we would want out of this. Pausing this work to collect feedback about what has been done so far and where we want to go next.