* add pool2d grad grad

* dbg

* add unittest

* update format

* add more unittests

* dbg

* Optimized the adaptive_avg_pool2d op when output_size == 1

* pool2d cuda kernel supports fp16

* fix compile issue of template

* add ut

There are mainly following fixes:

- take `DeviceContext` as the template parameter of math functors and OpKernel instead of `Place`
- remove `eigen_device` interface in base class  `DeviceContext`
- remove `GetEigenDevice` interface in `ExecutionContext` and base class `DeviceContext`
- remove unused `platform::EigenDeviceConverter`
- rename `REGISTER_OP_GPU_KERNEL` to `REGISTER_OP_CUDA_KERNEL`
- rename `USE_GPU_ONLY_OP` to `USE_CUDA_ONLY_OP`

* init softmax grad op

* add compute code

* export Backward to python

* update test ,export op.type to python

* update python test, fix compute bug

* update unit test

* use eigen

* optimize eigen code

* add gpu test

* register softmax_grad GPU kernel and fix test bug

* typo

* follow comments

Make implement an operator less noisy.

* Also fix unit test

* Implement InferShape and register them, give a stub Kernel method
  by LOG(INFO)