Created by: tonyyang-svail

A simple implementation parallel_do supporting multigpu

...
|
ParallelDo
    |		Split input
    |		Copy parameter to multiple GPUs
    -		Wait
    ||||	Forward on multiple GPUs
    -		Wait
    |		Merge output
    -		Wait
|
...
|
ParallelGradDo
    |		Split output@grad
    -		Wait
    ||||	Backward on multiple GPUs
    -		Wait
    |		AllReduce parameters
    -		Wait
|
...

TODO:

• Correctness
• init device context https://github.com/PaddlePaddle/Paddle/pull/7345
• getPlaceOp https://github.com/PaddlePaddle/Paddle/pull/6732