Created by: chengduoZH

fix https://github.com/PaddlePaddle/Paddle/issues/4637

• cudnn_helper.h can be used to process 3D data directly. I only add kNCDHW element and change it's unit test. class DataLayout are not used in current code.
• I only add register code for pool3d_cudnn_op and unit test.
• Although there are two interface for tensor descriptor in cudnn (cudnnSetTensorNdDescriptor and cudnnSetTensor4dDescriptor), but the current understanding is that cudnnSetTensorNdDescriptor can replace cudnnSetTensor4dDescriptor, and because it is only a Descriptor, so it is possible that the efficiency of influence is not large (or even negligible).

I'll get a better understanding of the difference between them. If someone know that, please tell me, thx.