Fix number of samples for no_trainer scripts

What does this add?

This PR fixes all of the no_trainer scripts to properly use the right number of training steps after the length of the dataloader was changed with accelerator.prepare

Why is it needed?

Currently in a multi-process setup, the progress bar still shows the old number of samples. As a result the old number of steps before breaking is set at the original amount, even though the length of the dataloaders changed. The progress bar reflects this too.

Simplified example:

If the dataloader starts with 128 batches, if 2 GPUs are used then each dataloader has 64 batches. As a result the progress bar should use 64, and the break condition needs to also know there is only 64. Both currently use 128 still

What parts of the API does this impact?

User-facing:

All scripts have a recalculation of the max_train_steps after accelerate.prepare

Basic Usage Example(s):

    # Prepare everything with our `accelerator`.
    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
        model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
    )

    # We need to recalculate our total training steps
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch

When would I use it, and when wouldn't I?

While this is always used, technically it is only needed when the number of nodes > 1.