[RLlib] Add 2D box example for PPO RL Modules

[ArturNiederfahrenhorst]

Why are these changes needed?

Historically, RLlib has simply flattened 2D inputs.
This often makes it hard for algorithms to learn in such spaces.
The Catalog introduced with RLModules throws an error when facing such spaces instead of silently flattening them and tells users to transform observations into a 3D space.
This example shows how this can be done on a pettingzoo env.

[ArturNiederfahrenhorst]

Executing the example leads to the following learning curve:

https://tensorboard.dev/experiment/xlxnhV0jQBqK13iRJvRzTA/#scalars

[kouroshHakha]

Is this not the same as what we have in the unittest?

[ArturNiederfahrenhorst]

@kouroshHakha Yes, we have unittests that cover these inputs.
But we don't have an example that shows how, after RLModules, users should interface with environments that have a 2D obs space. The solution here is simply to wrap the environment and transform the observation space into something that RLlib likes. This is a simple solution that lets users write python code to solve their problem instead of implicitly transforming their inputs which has caused a lot of confusion in the past.