Make sky gpunode reuse existing cluster if possible

[ewzeng]

Partially fixes #1331 by making sky gpunode reuse existing cluster if possible.

Before:

$ sky gpunode --gpus V100
$ sky gpunode --gpus V100 # fails

After

$ sky gpunode --gpus V100
$ sky gpunode --gpus V100 # reuses gpunode

Tested

• sky gpunode --gpus V100 followed by sky gpunode
• sky gpunode --gpus V100 followed by sky gpunode --gpus V100
• sky gpunode followed by sky gpunode --gpus V100 (should error)
• The above tests, with -c name

Notes:

• I wanted to add a smoke test, but I couldn't because sky gpunode automatically ssh's into the cluster and there is no detach option.
• It is not possible to check if sky gpunode was launched with the exact same resource specification (since some resources are decided during execution), so we reuse the cluster if the required resources are a subset of the launched resources.

[romilbhardwaj]

Thanks for the quick fix @ewzeng!

Just a small note after running some tests - sky gpunode –gpus V100 followed by sky gpunode correctly logs you in to the gpunode on current master (464b5db). It's only the sky gpunode –gpus V100 followed by sky gpunode –gpus V100 case that results in an error. Before merging, please update the "before" part in the PR description for future reference:

$ sky gpunode --gpus V100
$ sky gpunode --gpus V100 # fails
$ sky gpunode  # reuses gpunode

[ewzeng]

Ah -- my bad. Thanks for catching that!