make fetch variable's shape consistent with output of inference. support numpy array input for predict interface

1) To build paddle-serving-server
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ -DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so -DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python -DSERVER=ON ..

2) To build paddle-serving-client
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ -DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so -DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python -DCLIENT=ON ..

3) To build paddle-serving-app
cmake -DPYTHON_INCLUDE_DIR=$PYTHONROOT/include/python2.7/ -DPYTHON_LIBRARIES=$PYTHONROOT/lib/libpython2.7.so -DPYTHON_EXECUTABLE=$PYTHONROOT/bin/python -DAPP=ON ..

fix general op reader problem, make fetch var changable during inference; add more arugments for resource

make general_model_config unified on client side and server side
make general_model_config accessible with Python API