Optimize the CUDA implementation of sequence_expand op by reduce the times of copying lod data from CPU to GPU. (#15493)

* Optimize the CUDA implementation of sequence_expand op by reduce the times of copying lod data from CPU to GPU.
test=develop

* Refine the op benchmark to support setting lod in config.
test=develop

* wip simplify operator framework

* wip

* wip

* done test=develop

* clean test=develop

* fix test=develop

* fix deps test=develop

* fix cpu build test=develop

* fix tensorrt build test=develop

* fix tests test=develop

* fix test=develop

* fix cpu build test=develop

* "re-commit "

* "picked up"

* "fix ci"

* "fix pdb hang up issue in cuda 9"

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`

Make implement an operator less noisy.

* Implement InferShape and register them, give a stub Kernel method
  by LOG(INFO)