Support to create a gradient machine with merged model in c-api

related codes in Paddle core.

to create a gradient machine for inference using merged model with parameters
which is genearted by `paddle merge_model`.

Feature/fix random op link error

support python test without installation python package

Fix the bug between nvcc and boost

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

turn off with_mklml and with_mkldnn to fix ImportError of libmklml_intel.so

Fix #3386

 tensor slight improve

fix test_KmaxSeqScore for only CPU compile.

gather function only

Random op

A better error message for gradient checker

Remove unnecessary C++ operator test