[Feature Request] Create an optimizer hook for the trainer

[vmoens]

Motivation

Optimizers are currently passed as a single object to the optimizer. In some cases, we'd like to have multiple different optimizers for a problem (for instance to update the parameters only every N operations).

Optimizers should be implemented using a hook (for instance _optimizer_hook placed in between _post_loss_hook and _post_optim_hook, and that would receive the loss as input) that would call each optimizer.

Since the loss is a TensorDict of losses, we could even contemplate the possibility of having a backward for each optim hook, and hence each optim hook would do its own backward, optim step and zero_grad over its set of parameters.

We should keep the optimizer in the args of the Trainer.__init__ method as it's convenient for the users to pass it there, but optimizer=None should be supported (then the _optimizer_hook would not see anything and would just be a no-op).

Doc: https://pytorch.org/rl/reference/trainers.html#torchrl-trainers-package

[vmoens]

Solved by #716