Created by: QiJune

• Refine cuDNN related operators:
  • rename GPUPlace to CUDAPlace and introduce CUDNNPlace derived from CUDAPlace https://github.com/PaddlePaddle/Paddle/pull/6518
  • introduce CUDNNDeviceContext derived from CUDADeviceContext https://github.com/PaddlePaddle/Paddle/pull/6585
  • make current cuDNN related operators to be a OpKernel of corresponding operator. For example: conv_cudnn is actually a OpKernel of conv operator, it should not be a independent operator. https://github.com/PaddlePaddle/Paddle/pull/6660
• Refine the interface of OperatorBase::Run, take vector<Place> as parameter instead of DeviceContext https://github.com/PaddlePaddle/Paddle/pull/6783
• Implement a pool of DeviceContext in init stage https://github.com/PaddlePaddle/Paddle/pull/6537
• refine operator run logic. https://github.com/PaddlePaddle/Paddle/pull/6783