Add training and inference modes for the envs

[victor-villar]

There are things done during training that are not needed during inference (reward calculations in envs for example). It might be good to add a training/inference mode for the envs to enable/disable the appropriate things.