Feat: deploy MPI-Operator Standalone

[Souheil-Yazji]

Is your feature request related to a problem? Please link issue ticket

The training operator suffers from many issues which impact performance, scheduling and other factors. The training operator's bundled mpijob crd is also out of sync with the MPI operator CRD for KF1.7 manifests.

Describe the solution you'd like

Implement MPI Operator standalone and allow it to operate without removing the training operator deployment

Additional context

kubeflow/trainer#1972

[Souheil-Yazji]

First, the training operator deployment was modified to add the enable-schemas args, which prevent the training operator from watching mpijobs by default.

Second, the mpi operator was deployed using kustomize pointing to the v0.5.0 tag. The CRD for mpijobs was manually updated using kubectl replace since updating the last applied config annotation causes kubectl apply, which is used with kustomize, to break due to exceed file size limit:
The CustomResourceDefinition "mpijobs.kubeflow.org" is invalid: metadata.annotations: Too long: must have at most 262144 bytes