Create a simple, reproducible, out-of-the-box snapshot for doc classification paper

[daemon]

The document classification paper deserves a clean snapshot for reproducibility and extensibility. Readers should be able to click the link in the paper, run a few commands in the README, and reproduce the results in no time. AWD-LSTM is extremely popular (and highly cited) for this reason, as well as Honk on a lesser scale.

[achyudh]

As suggested in the meeting today, let's split Castor and deal with cleaning up code for the single text sequence tasks first. It would be nice if we can have independent issues identifying all the problems we currently face with Castor