Fit Google C++ Style

* 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.

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

Fit google C++ style

* 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

* insert rnn's backward into Backward()

* add device_context into backward_test

* make stepnet member

* add pybind support

* fix Inputs Outputs

* remove unique_ptr

This reverts commit fb61512c.

* Should always use `dynload::` for cuda function.
* Fix cublas.h without DSO load.