Reorganize into three python packages: toybox-cffi, toybox-api, and toybox-envs

[jjfiv]

Fixes #139, #95.

The idea is that:

• toybox-cffi package is pure-rust, compiled into a wheel, now using pyo3-pack command, rather than the semi-unmaintained milksnake.
• toybox-api provides Toybox, Simulator, and State types as well as intervention API.
• toybox_envs depends on gym and provides atari Environments. This can be versioned separately for when gym updates their API and we don't need to change anything lower. (currently doesn't function with the latest gym because they changed how register works), but that's intertwined with our use of baselines for experiments. We ought to be able to remove baselines (#78) now and put experiment code in separate repositories (per paper?).

Since this moved a bunch of python code around, everyone's a reviewer. To get the envs loaded, now we call:

import toybox_envs

As:

import toybox

Will only get you to the interventions/Toybox/Simulator/etc. apis.

I found some tests we weren't running (the main methods in the intervention python scripts) and added them to travis.

./start_python from ctoybox still works for development/training as long as you're in a virtualenv.

[jjfiv]

Also, if you make pypi accounts (and verify your emails @etosch) I will add you both as owners to the packages, but I'm happy to take the lead on deploying versions for now since it's a pain.

[jjfiv]

This is superseded by #149.