* "cherry picked operators changes"

* "remove duplicated code"

* "add constant setter"

* "add get expected kernel"

* "fix ci"

* "add fill constant"

* "add gradient register"

* "make some enhance"

* "better format"

* "fix typo"

* "fix reuse"

* "fix get expected kernel"

* "change the mkldnn code"

* "fix mkldnn"

* "fix mkldnn failed test"

* "add comment"

* fix __shfl_down_sync_ of cross_entropy

* use reduceSum

* "fix ci"

* "re-commit "

* "picked up"

* "fix ci"

* "fix pdb hang up issue in cuda 9"

* add cuda_device_functions.h

* move reduceSum to elementwise_op_function.h

When the second argument contains batch dimension, the axis should be 0.

Also makes elementwise ops more tolerant at handling tensors with trailing
singular dimensions.

* Speed up elemwise grad

* Fix bug

* Add macro for MAX_BLOCK_DIM

And some minor fixes on comments.

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`

* Fix Proformance problem of enforce

* Fix missing `;` in code

* Fix CI