fix(data_loader): handle dataloaders missing batch_sampler attribute

[SparshM8]

What does this PR do?

This PR fixes issue #882 where prepare_data_loader (and consequently Accelerator.prepare) fails with an AttributeError when a DataLoader object does not have a batch_sampler attribute. While standard PyTorch DataLoader objects always have this attribute, some custom or wrapped dataloaders might not, causing crashes in accelerate.

Changes

• Updated get_sampler helper to safely access batch_sampler using getattr.
• Updated prepare_data_loader to use getattr(dataloader, "batch_sampler", None) instead of direct access.
• Added a native regression test tests/test_issue_882.py that simulates a dataloader without a batch_sampler attribute.

Validation

• Verified with a reproduction script that was previously failing.
• Native regression test passes locally.
• Confirmed that it still preserves custom sampler structures if present.

Fixes #882