Homogenise and debug RL solvers and implement action masking in RLlib

[fteicht]

This PR both improves the RL experience in scikit-decide and corrects some bugs in the RL solvers:

• Implement action masking in the RLlib solver for the domains that implement _get_application_actions_from()
• Implements the bi-directional conversion from numpy arrays to user-defined classes in the scikit-decide spaces deriving from gym.Space
• Allows DictSpace and TupleSpace to use subspaces optionally deriving from GymSpace, and implements proper recursive unwrapping and wrapping of elements belonging to those subspaces
• Implements a SetSpace space that derives from GymSpace and that provides efficient search in the contains method
• Allows the hub domains Maze, SimpleGridWorld and Mastermind to be solvable by RLlib in addition to StableBaselines
• Correct a bug in scikit-decide's gym domain where the reset method did not unwrap the initial observation
• Correct a bug in scikit-decide's StableBaselines solver where the _sample_action method did not unwrap the current observation
• Improve coverage of RL solvers in the unit tests

[neo-alex]

Great job, thank you!