Universal Operator Phase I: Merge apis and controllers from all repos to all-in-one-operator repo

[Jeffwan]

Part of kubeflow/common#138
In Phase I, we want to merge different operators's apis and controller in one place.

• Tensorflow http://github.com/kubeflow/tf-operator
• PyTorch http://github.com/kubeflow/pytorch-operator
• MXNet http://github.com/kubeflow/mxnet-operator
• XGBoost http://github.com/kubeflow/xgboost-operator

MPI is not part of it, Let's consider it in later stage since the distributed strategy is totally different from rest of them and implementation are not sharable

[Jeffwan]

XGBoost - #1286 #1293
PyTorch - #1294
Tensorflow - #1295
MXNet - #1296

[Jeffwan]

The goal of Phase I is to make code compilation work in this branch. controller can run with simple example. No need to cover all the tests.

[Jeffwan]

We have all related PRs merged. This can be closed.