Develop new learning algorithm: PPO

[nick-harder]

We should start working on a new DRL algorithm based on MA PPO algorithm, it promises significant speed improvements, and would solver the critique of the centralized critic approach

[nick-harder]

This tasks has been given low priority as other issues need to be adressed first

[kim-mskw]

The general structure is ready on the PPO branch and runnable with one gradient step. However, the conversion for a single agent seems to get stuck in extreme values, so nothing too valuable is learned.