[SDK] Create unify client for all Training Job types

[andreyvelich]

Recently we introduced SDKs for MPIJob: #1608 and XGBoostJob: #1607 in a separate Kubernetes Python clients.

All of our clients (TFJobClient, PyTorchJobClient, MPIJobClient, XGBoostJobClient, MXJobClient) use logically the same CRUD APIs and we can reduce our code base.
Also, from the user perspective I don't think it is neseccary to create various Kubernetes Python clients for each training framework (e.g. TF, PyTorch, etc.)

I propose to create the unify TrainingClient (similar to our unify Training Controller) and to define APIs for each job under this class. For example, create_tfjob, create_pytorchjob, create_mxjob_from_func, etc.
I think that will simplify our UX and SDK maintenance effort.

/kind discussion
@kubeflow/wg-training-leads @tenzen-y @anencore94 What do you think ?

[andreyvelich]

/area sdk

[terrytangyuan]

Great idea!

[johnugeorge]

Agree. It will give a better user experience

[tenzen-y]

Looks great!

[andreyvelich]

/assign @andreyvelich