Support openai.gym interface

[kantneel]

gym is the most commonly supported environment interface. People who have been experimenting with agent designs may have already built around the gym interface i.e. obs, reward, done, info = env.step(action).

There is some old code in env/src/gym which appears to have once been a gym interface for FLE. It seems out of date though since the action space and observation space do not correspond to the python REPL that FLE uses otherwise. It would be great to update this so that people who are used to gym interfaces can plug into FLE more seamlessly.