Created by: typhoonzero

Fix https://github.com/PaddlePaddle/Paddle/issues/4194 Fix https://github.com/PaddlePaddle/Paddle/issues/4187

TODO list all done, please review, thanks~

• Reuse CPU code after https://github.com/PaddlePaddle/Paddle/pull/4042 merged.
• Use paddle::memory to allock cudnn workspace.
• Support group.