Decoupled support for Keras-backend interface

[bvanessen]

Add the capabilities in LBANN to act as a backend for the keras framework. This would probably look something like:

0. Creating a LBANN prototext model based on Keras model description
1. Allowing the Keras train call to launch LBANN in a separate process
2. Somehow return control back to Keras

Some of the caveats are that we will only support layers that have a direct correspondence in the LBANN implementation. Can we expose an API into LBANN that allows Keras to reach into using a polling approach to monitor progress.

Note that this type of interface would not be a full implementation of the backend API, but it would allow LBANN to take advantage of network generation techniques in python.