Fix sparse update memory error for distributed training

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`

* Changing learning rate from attribute to input(float)
* Removing obsolete code

Fix #3902

* use operator context

* optimize code

* update net infershape

* update InferShape

* disable override InferShape(scope) in OperatorBase

* change InferShapeImpl to InferShape

* add template to OperatorContext Input/Output

* merge Input InputVar, Output OutputVar

* change Inputs to MultiInput

* fix conflict

* fix MultiInput bugs and add unit test

* rename KernelContext to ExecutionContext

* clean code

* change InferShape to protected

* fix template bug

* refine code

* use InputVar instead of Input<Variable>

* typo

* optimize code

Make implement an operator less noisy.

* a simplest SGD op

* add inputs

* add ut for multiple inputs

* fix AddToLayer

* op_desc -> op_proto

* CreateArgumentOffsetMap -> CreateInOutOffsetMap

* move CreateInOutOffsetMap from OperatorBase to op registry

* arg_idxs_ -> in_out_idxs_

* Convert `op` --> `operators`
* Remove AddType in OpProtoMaker, because type is part of registry.
* Rename CPU_OR_GPU --> DEVICE_TYPE in registry macro.

* Refine register methods, make Op can get rid of whole-archieve
* `USE_OP` before a op is used.
* Add unittest for add_op.