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

Adapting to the BatchNorm structure to support 3D data

Adaptive data structure for switch order layer

fix batch_norm parameter share