[Feature] Load dataset distributedly in the single-controller mode

[garrett4wade]

Checklist

• This feature will maintain backward compatibility with the current APIs in
  areal/api/. If not, please raise a refactor issue first.

Background

Loading data in the centralized controller process is extremely expensive.

Potential Solution

Create a DataController and distributed workers for loading and processing data in multiple processes. The worker processes will send RTensors back to the controller. The real tensor data will transfer between the underlying workers.

[guozhihao-224]

The difference between this issue and #907 PCLEVR Processor Perf is that one is vertical optimization and the other is horizontal scaling, right?

[garrett4wade]

@guozhihao-224 Yes. #907 focuses on optimizing the image processor itself, whereas this issue addresses an architectural change in where the image processing job is executed.

[garrett4wade]

Closed by #1120