* Some improvements to support bert mixed precision training

test=develop

* Revert the cast in layer_norm

test=develop

The macro should be defined by compiler rather than by source.

test=develop

* Revert "accelerate dropout (#9902)"

This reverts commit 2e331c65.

* Correct discard

* accelerate dropout

* accelerate dropout

* "fix the dropout test"

* "rerun ci"

* "fix ci"

* "rerun ci"

* "fix ci"

* "fix"

* "stage"

* disable

* "merge random generator kernel and mul"

* "fix dropout"

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`

* is_training to is_test in dropout op

* handle dropout and batch_norm operator when prune pdesc in testing mode

* handle dropout and batch_norm operator when prune pdesc in testing mode

* add get_inference_program method

* fix dropout op

* fix ci

* test data after each batch training

* refine code

* refine test_book3

* fix ci

* follow comments

Change type of dropout_prob to template typename.

Make implement an operator less noisy.

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