* clip softmax for fp16

* updates

* fuse xent support fp16 test=develop

* wip

* wip

* add simple row reduce

* wip fp16 accurate softmax

* add accurate softmax kernel for fp16 test=develop

* update test=develop

* fix cpu build test=develop

* update api.spec test=develop

* follow comments test=develop

* fix build test=develop

* fix trt build test=develop

* fix inference build test=develop

* fix merge test=develop

* update test=develop

* try fix build test=develop

* fix build test=develop

* rename real_exp test=develop

* fortest

* remove hacky kernels test=develop

* clean up test=develop

test=develop

test=develop

* "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`

OLD: op_type -> grad_op_creator

NEW: grad_op_type -> grad_op_creator
     op_type -> grad_op_type