Correctly use host_vector in LoDTensor and expose LoDTensor to Python.

Make paddle.v2.inference can direct load protobuf

refine all operator python tests using new test framework

fix DeConv3D switch(imgSize*_, output*_)

fix switchOrderLayer

fix attr bug in op_test and ensure order in duplicate inputs/outputs

Fix Op unit test failed

Refine MKLDNNMatrix and MKLDNNLayer

add concat op with CPU kernel

Remove enforce demangle

* get output when training

* follow comments

make docker can configure with_mkldnn and with_mklml

Fix switch layer attribute name

It is buggy in some Linux because the unique_ptr will be free however
the std::string trying to use that char*.

Moreover, it is no need to demangle for error log by Paddle.
Just use `c++filt` or other shell utilities to do this.

Fixed some ssd bugs.

Refactor operator python test framework and add sum operator