It can be run both CPU/GPU. configure attributes are:

* min: the min value of uniform random
* max: the max value of uniform random
* dims: the dimension of output tensor
* seed: the random seed of uniform random. 0 means generate a seed each
        time.

* init cross entropy graident

* add crossentropy grad op

* remove details

* fix static compile

Make ci pass

Currently use `Operator("fc", X="x", W='w1', B='b1')` as operator
creation method.

Fix #3198

Rename operator.py to op.py because it is conflict with protobuf

* init softmax grad op

* add compute code

* export Backward to python

* update test ,export op.type to python

* update python test, fix compute bug

* update unit test

* use eigen

* optimize eigen code

* add gpu test

* register softmax_grad GPU kernel and fix test bug

* typo

* follow comments

* export Backward to python

Add get_numeric_gradient API and its unittest.

* Make interface of Operator to `const Scope&`

* add rnn op interfaces

* add Run

* rename state -> memory

* change state -> memory

* make compilable

* add .cc

* init test

* add op fake implementation

* add CreateStepNet and CreateScopes implementation.

* add TODO list

* init memory attributes.

* add LinkMemories

* add PlainNet fake implementation

* Use std::shared_ptr<Scope> in the OpRunContext.

* add test

* disable mutable_data

* finist segmentInput function

* enable mutable_data with a trick

* RNNOp test.

* enable LinkMemories with mutable_data

* update SegmentInput function with comments

* finish ConcatOutput function

* reformat inputs and attributes

boot_memories

* Refine unit test.

* Refine unit test.

* modify inlinks.

* add OpDesc to Net

* fix bug and update unit test.

* move step scopes from inputs to outputs

* fix merge conflict, update SegmentInput function

* add RecurrentOpProtoAndCheckerMaker.

* clean the codes

* Abstract GetStepScopes and GetMaxSeqLen function

* refine LinkMemories

* Refine code and add some comments.

* add backward core

* update for develop branch.

* add forward core

* add forward algorithm

* Add RecurrentGradientAlgorithm implenmention.

* use CopyFrom and Slice function in RecurrentOp

* add unit test for LinkMemories.

* fix unit test.

* use the latest tensor.h, solve conflict

* add maker

* move SegmentInput and ConcatOutput to details nameplace

* unit test for RecurrentGradientAlgorithm.

* apply OperatorBase

* apply net operator.

* move memorys to attributes

* add RecurrentGradientOp

* open test unit test in recurrent_network_op_test.

* revert some files.

* add RecurrentArgument and Link struct to simplify member variable.

* rename.

* move recurrent_op from framework to operators

* add RecurrentGradientOp Init

* fix name

* fix Link.interal/external name

* use namespace operators instead of framework

* clean the code

* use the latest add_op and mul_op, don't test backward now

* Remove ScopePtr and OperatorPtr

* add get_net to pybind

* add test_recurrent_op.py

* add random into gen_tensor

* update to develop branch and refine some code.

* add some comments.

* Refine unittest also

* Expose Net to Python

* Expose PlainNet to Python, make python can add_op, complete_add_op
* Provide a low level api to manipulate Net
* Unittest for Net::DebugString

* add cross-entropy-op

* add infershape and compute

* implement Infershape and compute of onehotcrossentropy op

* a simplest SGD op

* Expose PlainNet to Python, make python can add_op, complete_add_op
* Provide a low level api to manipulate Net
* Unittest for Net::DebugString

* Fc operator is a grouped operator, which combined by may internal
  operators.
* InferShape & Run a FC operator in Python.

* Follow review comments to seperate Tensor Numpy interactive methods in
  tensor.h.
* Simplify logic for `CastToPyBufferImpl`, make it as one struct and in
  details namespace.
* Remove `Scope` expose in Python, since it currently is useless.
* Remove some debug functions.

Basically following
http://pybind11.readthedocs.io/en/stable/advanced/pycpp/numpy.html

* Use buffer protocol to return a view of Tensor. It can be cast to
  numpy array in Python.
* Set a numpy array to a tensor.

All OpCreation method are generated by
`create_op_creation_methods::__bootstrap__` method, and stores in
`op_creations` object and its methods.

There are three parts to implement this feature.

1. Get all registered `OpProto` from C++ side. It is implemented in
`get_all_op_protos` method.
1. Create a function to convert `kwargs` to `OpDesc` base on each op's
`OpProto`. The `OpDescCreationMethod` class.
1. Convert `OpProto` to `docstring` by `get_docstring_from_op_proto`
method.

All three methods are unit tested. The `__bootstrap__` just combines
them together and create a method in runtime.

For details, please reference the doc string in
`create_op_creation_methods.py` and the unit test
`test_op_creation_methods.py`.

* PyBind and SWIG of paddle cannot be load in a single Python process,
  lazy import all SWIG library of Paddle. Otherwise, the glog, gflags
  are imported twice in a same Python process.
* Note that all PyBind11 return C++ std::string as an unicode. For
  protobuf, it is need be cast to `str` before use them.
* Add unit test for Get `OpProtos`

`Paddle` manages Scope as programming language's scope.  It just a
thread-local stack of Scope. Top of that stack is current scope, the
bottom of that stack is all scopes' parent.

Invoking `create_var/get_var`  can `create/get` variable in current
scope. Invoking `enter_local_scope/leave_local_scope` can create or
destroy local scope.

A `scoped_function` will take a `function` as input. That function will
be invoked in a new local scope.