Debugging PPO training

[jcallaham]

Explanation

Previously the cylinder environment would train with the PPO implementation in RLlib but the performance would not improve (fixes #49). I think there were a few issues, including

• Environments (and their flows/timesteppers) not properly resetting
• Control inputs not being properly applied within rollouts

These at least seem to be fixed, although now there's a new issue where RLlib crashes after 3 epochs with a hard-to-understand error message (see #55).

Current status

• All tests passing
• RLlib will run without error for some amount of time and then crash

Plan

The way the package is currently structured makes the memory issue a bit hard to debug, especially with recent version changes in OpenAI Gym. Even though our package is in a sense "broken" at this point with respect to training RL agents, it's no more broken than it was before this PR. Since fixing this issue may require resolving #54, #55, and #56, I plan to merge as-is and then try to work through all these with a cleaner package (post-#56).

[jcallaham]

Also fixes #55