Add MultiBoxSpace handling in addReward

[carlocagnetta]

Trying to run an experiment with LabelmapSliceAsChannelsObservation, with FrameStack and AddRewardDetails,

• The experiment with DQN stopped during data collection:
  Process finished with exit code 137 (interrupted by signal 9:SIGKILL)
• The experiment with ActionRewardObservation also did not work:
  TypeError: Object -0.5694444179534912 in Batch( add_reward: -0.5694444, reward: -0.5694444, action: 0.0, add_action: 0.0, ) has no len()

[carlocagnetta]

@MischaPanch I tried to re-implement the DictObs without flattening, mainly because I would like to add an experiment with LabelmapSliceAsChannelsObservation but it did not work.
If you want to try it out, it would be great to add this to the experiments.