Bandits are non-TorchScript-able in 3.10

[tcleberg]

Hello- applying a bandit from Pearl and I came across a torch.jit.save() issue for a bandit class in a python 3.10, torch 2.0.1 environment:

This is due to two issues:

1. The more trivial one - the current type hints are not 3.10-compatible. We could very likely handle that by:

from __future__ import annotations

2. Optimizers are attributes of the policy learner (and the corresponding PyTorch type hints are not compatible).

Externalizing the optimizer from the policy_learner might make sense in any case.

        self._optimizer: torch.optim.Optimizer = optim.AdamW(
            self.model.parameters(), lr=learning_rate, amsgrad=True
        )

I'm aware that this is quite a restricted environment and that the solve might be to pin Pearl at versions above that - but wanted to log the issue.