Add gpu monitoring

[SandersAaronD]

Have the python exporter push gpu metrics upstream to mimir.

We can use imports from the (nvidia-maintained DCGM exporter)[https://github.com/NVIDIA/dcgm-exporter/tree/main] as appropriate but that is not necessary at this stage. The DCGM exporter is, among other things, fundamentally designed to be deployed into a cluster, and parts of its design may not work well for our use case without substantial work.

As a fallback implementation, we can run
nvidia-smi pmon
or
nvidia-smi dmon

to get good metrics for both processes and devices, respectively, and parse those out into prometheus metrics.