* add_op could take a unique_ptr or a const reference. If unique_ptr is
  taken, the NetOp will take care of that operator's life cycle. If a
  const reference is taken, that op will be Cloned.

Move OpRegistry functions to .cc file and move `OpMaker` to module operator

* Also simplify pybind implementation by using OperatorBase as holder
  type.

Fix sgd trainer document error.

Add a c-api interface to get the output of a specified layer.

Fix bug of enforce when dladdr not found

* Wrong Printf format before

add shared warpctc lib in whl

Fix invalid paddle executable file path with pip install  

Fit google C++ style

Make code more clearer

Implement EigenCudaStreamDevice

use param header to save mkldnn format info

* Clone method will create a new object instance, which is as same as
  itself.
* This is the first step to remove shared_ptr for OperatorBase

* add clang-format with version check 3.8

* improve doc

* insert rnn's backward into Backward()

* add device_context into backward_test

* large model train note

* update

* update doc

fix crash when disable WITH_SWIG_PY