Trainable implementation of full autoencoder.

[cellistigs]

Summary

This PR should get us to a first iteration of a pretty general purpose autoencoder without too many bells and whistles.

What has been done so far:

Straightened out a bunch of implementation details since the last pull request, including a script which has integrated training and evaluation for single sessions, including:

• Logging and checkpointing.
• Additional configuration files for training and model parameters.
• Built in evaluation loop (currently only on held out validation data for stupid memory reasons).

Remaining todos:

To a first approximation, this implementation leads to a trainable model, which generates fairly realistic looking images. Before merging, I'd like to figure out the following in this PR:

• Sequence dimension #2: Understand how to manage the sequence dimension for the data.
• Other dataloading options #3: Efficient means of dataloading for large, trial separated videos.