[Feature Request] Create a categorical tensorspec and make DQN primitives compatible with non-one-hot encoding

Motivation

We opted for default one-hot encoding for categorical data because it facilitates the retrieval of action values:

chosen_action_value = action * action_values

However, this is not scalable to large action spaces.
We should make it possible to use integers instead.

To make this possible, the following changes are needed:

• Create a DiscreteTensorSpec similar to MultOneHotDiscreteTensorSpec + tests
• Change QValueHook to make it possible to use a "categorical" space (or similar) instead of "one_hot"
• Do the same with DistributionalQValueHook
• Adapt the DQNLoss to these changes
• Adapt the DistributionalDQNLoss to these changes
• Change the gym specs reader here to make it possible to choose one spec type or the other. One way to go about this would be to have a global variable that controls which categorical encoding must be used

  ONE_HOT_ENCODING = os.environ.get("ONE_HOT_ENCODING", True)
  

  Then we can use this in all the code-base to choose one or the other encoding.

All these features should be unit-tested and parametrisable (e.g. with ONE_HOT_ENCODING) in the DQN training scripts.