Refactor the workflow notebook

[wlandau]

I have enjoyed diving into your drake-powered workflow, and I am learning a lot about recipes and keras in the process. In this PR, I propose a modified workflow that trains multiple models and compares them in terms of performance. Advantages of the new R Markdown document:

1. It motivates the value that drake brings to a deep learning workflow. You can train multiple models in parallel (see the tips at the bottom) and skip models that are already up to date.
2. The new drake plan is easier to read, and it has more room for exploring multiple models, which is what plans were designed to do.
3. We duplicate less data in the new plan.
4. The new workflow serializes and unserializes the models according to keras best practices while preserving most of the convenience of drake's caching system. That way, make() can run post-processing on models trained in a previous session.

cc @edgararuiz

[wlandau]

By the way, this PR fixes #4 and #5.

[edgararuiz-zz]

Awesome work, thank you Will!