The rrllib wrapper should provide a single agent environment with `action_space` and `observation_space` attributes

[fteicht]

🐛 Bug

Many algorithms based on OpenAI gym domains expect the domain to define the action_space and observation_space attributes, and especially ray's rllib's single agent algorithms like AlphaZero.

See for instance ray's rllib's AlphaZero implementation which makes use of those attributes.

The current implementation of scikit-decide's rllib wrapper provides only a multi-agent environment wrapper via AsRLlibMultiAgentEnv which does not define the action_space and observation_space attributes (which is fine for rllib's multi-agent environments). Therefore scikit-decide's rllib wrapper should additionally provide a single-agent environment wrapper for algorithms like rllib's AlphaZero, which defines the action_space and observation_space attributes.

To Reproduce

Define a scikit-decide RL domain and pass it to ray's rllib's AlphaZero algorithm.
The following exception will be thrown when solving the domain:
AttributeError: 'AsRLlibMultiAgentEnv' object has no attribute 'action_space'

Expected behavior

No exception is thrown because an environment wrapper likeAsRLlibSingleAgentEnv (to be defined) should define the action_space and observation_space attributes.