Created by: dzhwinter

First part:

• Implement CUDNNDeviceContext, MKLDeviceContext hierarchy.
  Use decorator design pattern is not fit our needs so well, doing other things first.
• define LayoutType key #6827 (closed) #6832
• define LibraryType key #6770 (closed) #6874
• define the new OpKernelType class(four keys DataType/LayoutType/Place/LibraryType). #6769 (closed) #6879
• refine current kernel register mechanism
• refine CUDNN related operators, change them from operators to kernels https://github.com/PaddlePaddle/Paddle/pull/6660
• Remove CUDNNPlace and MKLDNNPlace follow new design
• rename GPUPlace to CUDAPlace
• add multikernel python test support

Second part:

• refine Tensor implementation, add a layout attribute #6765 (closed) https://github.com/PaddlePaddle/Paddle/pull/6955
• refine Python interface and some data operators to set layout.
• share Tensor layout in most operators

Third part:

• DataTransform function interface.
• DataTransformFn register mechanism #6823 (closed)
• kernel hint and GetExpectedKernelType #6883
• memory switch mechanism #6989 (closed) #6991
• refine memory switch mechanism in local scope #7057 (closed) #7058
• implement some basic DataTransformFn
  • CPU <---> CUDA #7050 (closed)
  • layout transform, like mkldnn
  • change batch norm to use new transform method
• add method to judge which inputs should be transformed refs
• Python interface setting kernel hint
• helper function of getting appropriate DeviceContext #7065 (closed) #7066
• Add a mnist example that switch kernel
• Refine current CRF operator