[API][discussion] Create a CRD for study

[gaocegege]

We have some discussions on slack and I have a idea to define a CRD and operator to manage studies, thus we could eliminate the customized CLI katib-cli and reuse kubectl to do the same thing.

I opened the issue to keep track of the idea.

[ddysher]

Two random thoughts:

• should we create CRD for trails as well?
• since hyper-tuning system is fairly standalone system, we might also consider using API aggregation, which gives you higher flexibility, e.g. we can keep using mysql as our storage backend but provides the same API interface to katib users

[gaocegege]

Yeah, I think so. API aggregation is more appropriate.

But it requires lots of work, and I am not sure if we have plan to do it.

[gaocegege]

I am not sure if we should create CRD for trails, since we define it in study config. Thus I think we could treat it as worker/ps in TensorFlow.

[jlewi]

Duplicate of #87