update dataloader wrappers to have total_batch_size attribute

[pacman100]

What does this PR do?

1. A requirement was to get resulting batch size post accelerator.prepare call. Adding total_batch_size to provide the resulting batch_size across devices.

my_dataloader = accelerator.prepare(my_dataloader)
accelerator.print(f"The resulting batch_size across devices is {my_dataloader.total_batch_size}")

[HuggingFaceDocBuilderDev]

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

[sgugger]

Sounds like a useful addition, but I'd prefer have this as a property and no new arguments in the init that aren't really used to influence the behavior of the class.

[sgugger]

Very clean!

[muellerzr]

Thanks for this! Seems quite useful indeed. My main concern here is that everything is documented properly given we're adding a new property to every dataloader type.

Worrying about the XLA wrapper in the docs is something we can discuss a bit later.

[muellerzr]

Great work!