Decouple prepare_data_loader() from Accelerator

[siddk]

What does this PR do?

Following #3039, decouples prepare_data_loader() (and all DataLoader classes/mixins in data_loader.py) from relying on AcceleratorState (and by extension, a properly initialized Accelerator), instead relying on PartialState. Fixes and extends tests in test_data_loader.py to validate both the non-Accelerator and decoupled workflows.

The benefit of this is that now folks can use prepare_data_loader and the nice features it provides (put_on_device, split_batches) without restructuring the rest of their training code around Accelerator.

Before submitting

• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@muellerzr @BenjaminBossan @SunMarc

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[muellerzr]

Very nicely done. I ran the tests locally and on two GPUs we do in fact pass! Nice work :)

[SunMarc]

Nice work @siddk !