Created by: reyoung

We should make Paddle fast even we using Operator. So we should give a way to fuse many fine-grained functors to a single GPU kernel.

The design doc should be written first.