Pass GPU diagnostics from worker to scheduler

[mrocklin]

This does a few things:

1. Use pynvml to collect information about any CUDA GPUs present
2. Optionally add those metrics to the worker's initial handshake and
   heartbeats
3. Collect that information in the scheduler in the WorkerState object

For now these just hang out in the scheduler information,
but in the future they might be used for dashboards,
or possibly scheduling decisions in the future.

I believe that everything gpu-specific here is fairly well separated
and generalized (others should be able to follow this pattern to add
more diagnostics relatively easily) but it would be good to hear from
others on if this is out of scope.

[TomAugspurger]

In general, this kind of special / extra information seems fine, as long as it doesn't affect the common case when the GPUs isn't present (and this one is fine).