Refine current codes to support multi-devices 2

[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 need to add layout type to fluid #6827 add data layout #6832
• define LibraryType key need to split concepts of Place and Library #6770 add library type #6874
• define the new OpKernelType class(four keys DataType/LayoutType/Place/LibraryType). need to refine our design of OpKernelType #6769 refine OpKernelType #6879
• refine current kernel register mechanism
• refine CUDNN related operators, change them from operators to kernels "cudnn operators change to cudnn kernel" #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 need to add a data member to represent the Layout of a Tensor #6765 "add data layout" #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 DataTransformFn register mechanism #6823
• kernel hint and GetExpectedKernelType Impl kernel hint #6883
• memory switch mechanism need to add memory switch mechanism in operator kernel switch #6989 add memory switch mechanism in operator kernel switch #6991
• refine memory switch mechanism in local scope need to refine memory switch mechanism in local scope #7057 cache memory in local scope #7058
• implement some basic DataTransformFn
  • CPU <---> CUDA implement CPU <---> CUDA DataTransformFn #7050
  • 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 need to get appropriate device context in switching kernel #7065 add helper function to get appropriate DeviceContext #7066
• Add a mnist example that switch kernel
• Refine current CRF operator

[dzhwinter]

This issue related with #6507