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`

* first commit

* Python API for while op

* Python Unittest for simple while_op forward

* fix out to be list

* Fix UT

* VarType

* Fix several bugs

* Fix bug

* Fix bug

* Fix Bug

* Fix bug

* Fix unittest

* Remove debug log

* Add comments

* add PADDLE_ENFORCE

* while_grad_op first commit

* Add `BlockDescBind::FindRecursiveOrCreateVar()` and fix bugs

* not sure how to setdim of while outputs

* push for test

* add executor vlog

* fix bug of while_op cond

* Several enhancement for code

1. Backward always infer shape & infer var type. Since there are RENAME
variables will be created when creating backward operator, but their
shape & var types are not inferenced.
2. Never use SomePtr-> directly, since every pointer could be nullptr if
it is a function return value. Add `detail::Ref` to cast pointer to
reference safely.
3. Enhance error message for backward.
4. Infer data type of variable in `sum` and `tensor_write`

* Fix bugs of while_op gradient

* Fix several bugs of while_op grad

* fix fill zeros like

* fix 3 >= 3

* fix place holder shouldn't be null

* fail on sum op

* Fix SumOp of TensorList

* clean up

* pass while test

* fix test_array_write_read

* pass sum op

* Support int/int64 for fill_constant_batch_size_like

* Fix compile

TensorSetConstant struct is used both in math_function.cc and math_function.cu. Somehow the release version can correctly handle it. But in debug version, set_constant_with_place() in math_function.cu uses the TensorSetConstant in math_function.cc and causes crash.

* initial matmul operator

Similar to np.matmul, but also has transpose_X and transpose_Y flags,
and only supports tensors from rank 1 to 3 inclusive.

For GPU, uses cublas?gemmStridedBatched. For CPU, uses
cblas_?gemm_batch if available via MKL; otherwise a simple serial
implementation that loops over the batch dimension is employed for now.

* Also use `const DeviceContext&` all the time, to prevent `const_cast`

Fix #4169
Fix #3468
Fix #3475