[k8s] CUDA envvars don't work in ssh

[romilbhardwaj]

CUDA-specific environment variables are dynamically set by the nvidia kubernetes operator and the node's kubelet when a command in a pod or kubectl exec -it ... is run. This allows the user to access the GPU in their python programs and through nvidia-smi.

Currently, these environment variables are currently not set when we run ssh mycluster. As a result, the GPU is unavailable when the user ssh's into the cluster for debugging.

This is related to #2287 and should be closed by #2435.