Refine current codes to support multi-devices

[QiJune]

• Refine cuDNN related operators:
  • rename GPUPlace to CUDAPlace and introduce CUDNNPlace derived from CUDAPlace Feature/multiplace #6518
  • introduce CUDNNDeviceContext derived from CUDADeviceContext "derived cudnnDevice context" #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. "cudnn operators change to cudnn kernel" #6660
• Refine the interface of OperatorBase::Run, take vector<Place> as parameter instead of DeviceContext Feature/operator run place #6783
• Implement a pool of DeviceContext in init stage Feature/global context #6537
• refine operator run logic. Feature/operator run place #6783

[wangkuiyi]

Let us use /~https://github.com/PaddlePaddle/Paddle/projects/35 to manage this list. After we create related TODO items in the project, we can close this issue. @QiJune