v1alpha2 pytorch API should try to be consistent with TFJob

[jlewi]

TFJob operator v1alpha2 made a bunch of API changes e.g

• Use dictionary for replicas not list
• Introduced CleanPodPolicy

We should create a v1alpha2 API for PyTorch that strives to be as consistent with TFJob as possible.

[jlewi]

@johnugeorge How's this coming? Looks like the PR kubeflow/training-operator#786 creates a v1alpha2 API for PyTorch that is consistent with TFJob v1alpha2.

[johnugeorge]

Yes. You are right. The confusion is only wrt to location where to keep the code. I tested v1 alpha2 code with a basic example.