Improve elementwise operators performance in same dimensions

add elementwise_sub_grad_grad op for backward of backward calculation

* 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

* fix input<tensor>
test=develop

* fix split_ids
test=develop

* ElementwiseMul should not support SelectedRows

* fix scale op
test=develop

* change GetTensorFromVar() method to GetTensorOrSelectedRowsFromVar()

* fix operator

* refine MultiOutput

* fix MultiOutput
test=develop

* disable test_dist_save_load
test=develop

* fix elementwise_op
test=develop

* add get_sparse_as_op
test=develop

* add info for check
test=develop

* rename get_sparse_as_op with extract_rows_as_op.
test=develop

* elementwise doesn't support selected_rows

* fix regularizer

* remove extract_rows_as
test=develop

* fix ci
test=develop

* add test for sum_op

* fix regularizer
test=develop

*  test=develop

* fix pserver weight decay multi inputs test=develop

* "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"

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`

Elementwise operator add/sub/mul/div