* relu forward opt

* add gelu functor

* optimize code

add approximation for gelu, default value is False (only kernel with eigen is added, remove code for computing gelu with MKLDNN temporarily)

* add erf op and python interface.

* add fp16 support for erf op.

* add unitests for erf op and its python interface.

* add selu

* use for range
test=develop

* add API
test=develop

* follow comment
test=develop

* update API.spec
test=develop

* Fix infer shape of crop op.
* Speed crop op.

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`

2. Fix unitest