Rein.predict

[polvalente]

We currently only have Rein.train, which trains agent models.
We should have Rein.predict that uses a given trained model for traversing and environment without training.
This means that the agent should skip it's training loop, and the neural networks, when applicable, should be used in prediction mode too.
Some agents will also need adaptations (e.g. SAC should have a flag for using just the mean action instead of the action distribution).