* "add softmax cudnn functor support"

* "add testing"

* "refine cmakelist"

* "sequence softmax forward speed up"

* "add softmax grad"

* "fix sequence softmax test"

* "add double precision'

* "fix softmax test"

* "add softmax cudnn support"

* "fix softmax cudnn test"

* "add softmax to nn.py"

* "fix compile bug"

* "refine cmakelist"

* "fix ci"

* "fix based on comment"

* "fix based on comments"

* "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`

* Remove template parameter for Tensor methods

* Also check the type is correct when data()
* Simplize holder_

* Fix accuracy_op

* Register Code

* No POD instance is forbidden in global scope.

See
https://google.github.io/styleguide/cppguide.html#Static_and_Global_Variables

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.