Re-define the way we build Docker images

Demo: Generative Adverserial Nets

Fix --job=test

When the define_py_data_sources2 has both train_list and test_list,
for job=test, the trainer will create both dataProvider_ and testDataProvider_.
But dataProvider_ is not used. This causes SIGSEGV at finishAsync() because asyncLoader_ is not created.

Change-Id: If579f715f80a70ebc795094792c3436bfa0f5746

refine build_docker_image.rst

Refine cmake version tag.

Unify conv

Added a test in test_NetworkCompare to verify cudnn and exconv, fixed a bug in cudnn_conv in dealing with groups

remove build of doxyfile

Using ApproximatelyEquals for trainer config unittest.