add support for auto-context in synthesis generator

[bodokaiser]

There are two possible strategies on how to proceed with this:

1. Extend rebuild.py to support an option which produces [depth, height, width, channels] tfrecords wherein channels corresponds to the i-th auto context iteration.

2. Update train.py to support loading various volumes and merge them via tf.contrib.Dataset.

For the moment I incline to the first strategy as it would

• distribute the implemented logic evenly between train.py and rebuild.py
• should provide better performance as there is less io involved

however we would need to adjust current apis to use DHWC over DHW.