Respect multi-GPU outputs in nvidia-smi

[charliermarsh]

Summary

This initially included NVIDIA_VISIBLE_DEVICES masking, though it's now omitted for simplicity.

Closes #14647.

[geofft]

This seems fine, but some nitpicks:

1. I don't actually think there's a point to us parsing NVIDIA_VISIBLE_DEVICES here. This variable is specifically used by nvidia-container-toolkit to determine which devices ought to be exposed inside the container. It doesn't seem to be something that's used in non-container tools at all. I think the reference to it in #14647 was just mentioning the lack of ability to use this as a workaround for being unable to parse multiple lines, but if we handle multiple lines I'm not sure we need the workaround. It doesn't really matter what we do here since the driver version ought to be the same for all lines of output, but if we extend this code to be about compute capabilities etc., I think we should put a tad more thought into whether we want this to be the interface, since I think we would be novel in using this environment variable in a non-container tool (e.g. I don't think that nvidia-variant-provider uses it).
2. Somewhat weirdly the parsing for the variable in nvidia-container-toolkit appears to allow all/none/void to be individual elements in the comma-separated list, as opposed to having to be the entire string, e.g., NVIDIA_VISIBLE_DEVICES=2,all is accepted and interpreted as all. See https://github.com/NVIDIA/nvidia-container-toolkit/blob/v1.17.8/internal/config/image/cuda_image.go#L123-L154 which splits on commas and passes a list to https://github.com/NVIDIA/nvidia-container-toolkit/blob/v1.17.8/internal/config/image/devices.go which loops through the list looking for these special keywords.

[charliermarsh]

Okay, sounds good. I removed the NVIDIA_VISIBLE_DEVICES parsing for now.