pytorch-operator: Consolidate manifests

[yanniszark]

Umbrella-Issue: kubeflow/manifests#1769

PyTorch Operator

Current manifests structure

.
├── crd.yaml
├── deployment.yaml
├── kustomization.yaml
├── namespace.yaml
├── podgroup.yaml
├── pytorch-job-crds
│   ├── base
│   └── overlays
│       └── application
├── pytorch-operator
│   ├── base
│   └── overlays
│       └── application

Explanation

• pytorch-job-crds and pytorch-job-operator came from kubeflow/manifests.
• The other files were the existing kustomization in the upstream repo.
• The application overlays contain Application CRs, which we want to move away from (Why do we need kubernetes-sigs/application controller in Kubeflow? manifests#1715 (comment))

Recommended end state

The main diff is that the upstream manifests also install a namespace resource, which we don't want when installing with kubeflow. Following the standard base and overlays structure, I propose the following structure:

├── base
├── overlays
│   ├── kubeflow
│   └── standalone

• overlays/standalone is used to install the operator in its own namespace, for testing or other purposes.
• overlays/kubeflow is used to install the operator as part of kubeflow.

[yanniszark]

mistakenly opened