[test] attempt to fix CI test for PT 2.0

[younesbelkada]

What does this PR do?

This PR fixes the failing tests related to Pytorch 2.0 upgrade

In Pytorch 2.0, the learning rate schedulers have been re-designed, thus we need a bit of refactoring on trl to deal with unexpected behaviors.

In Pytorch 2.0 it seems that the class _LRScheduler have been deprecated in favor of LRScheduler

In Pytorch 2.0:

# Including _LRScheduler for backwards compatibility
# Subclass instead of assign because we want __name__ of _LRScheduler to be _LRScheduler (assigning would make it LRScheduler).
class _LRScheduler(LRScheduler):
    pass

In Pytorch < 2.0:

class LambdaLR(_LRScheduler):
    """Sets the learning rate of each parameter group to the initial lr
    times a given function. When last_epoch=-1, sets initial lr as lr.

    Args:"""

Hence the check isinstance(lr_scheduler, torch.optim.lr_scheduler._LRScheduler) does not pass anymore with Pytorch >=2.0

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[younesbelkada]

In PT < 2.0 the schedulers are wrapped inside an attribute scheduler
EDIT: what I said is wrong, accelerate wraps it inside AccelerateScheduler only for python != 3.8

[younesbelkada]

Tracking the issue at: huggingface/accelerate#1200

[younesbelkada]

it is a way I have found to do versions checking wrt to python version, I would be curious to see if you know better techniques.