transfomer on a single device step time
    reduces from 0.157 to 0.125

* Update tensor_util.h

* Update with moved TensorDesc

* Fix tensur_utils.cu

* Update

* Update

* Update

* Update

* Make tensor_util.cu a symbolic link

* add data layout

* fix ci

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 batch norm op

* prepare input output

* compute mean_out var_out save_mean save_var on CPU

* active is test

* use eigen to do computation

* complete batch norm forward

* set default momentum to 0.9

* add batch norm grad op in CPU

* add tensor_format and NHWC support, add python test

* add test training

* add batch norm gradient test

* improve comment, fix foward Python UnitTest

* add gradient test

* fix eigen warning

* follow name style

* fix a bug

* change float to T

* add simple forward test

* test with different place

* add backward test

* refine python test

* remove old python test code

* code clean

* follow code style

* update comment

Make implement an operator less noisy.

* add inputs

* add ut for multiple inputs

* fix AddToLayer

* op_desc -> op_proto

* CreateArgumentOffsetMap -> CreateInOutOffsetMap

* move CreateInOutOffsetMap from OperatorBase to op registry

* arg_idxs_ -> in_out_idxs_

* Convert `op` --> `operators`
* Remove AddType in OpProtoMaker, because type is part of registry.
* Rename CPU_OR_GPU --> DEVICE_TYPE in registry macro.

* Refine register methods, make Op can get rid of whole-archieve
* `USE_OP` before a op is used.
* Add unittest for add_op.