Scope pytorch-operator to a namespace

[jlewi]

See kubeflow/kubeflow#1213
and kubeflow/training-operator#759

We'd like to be able to scope Kubeflow to a particular namespace to better support multi-tenant environments.

Right now pytorch uses a cluster role and claims jobs in all namespaces. We'd like to add an option to scope it to one or more namespaces.

Here's what we did for TFJob operator; kubeflow/tf-operator

Also worth considering what Argo does see
kubeflow/training-operator#759 (comment)