feat: add DispatchDataLoader for custom iterable dataloaders

[KrishVenky]

Closes #2975

Adds DispatchDataLoader, a lightweight wrapper that makes any Python iterable usable with Accelerator.prepare() without requiring it to be a torch.utils.data.DataLoader. Also adds custom_classes to DataLoaderConfiguration so the Accelerator can auto-detect and wrap user-defined iterable types.

What does this PR do?

Implements the feature requested in #2975 ("barebones dataloader for any iterable"):

• DispatchDataLoader (data_loader.py): wraps any iterable, calls __iter__ on each pass, and applies automatic device placement via send_to_device. Exported from the top-level accelerate package.
• DataLoaderConfiguration.custom_classes (utils/dataclasses.py): a tuple of class types. When set, Accelerator.prepare() detects instances of those classes and wraps them in DispatchDataLoader automatically.
• _prepare_one (accelerator.py): recognises DispatchDataLoader directly, and wraps custom_classes instances.

Usage:

from accelerate import Accelerator, DispatchDataLoader
from accelerate.utils import DataLoaderConfiguration

class MyDataSource:
    def __iter__(self):
        for i in range(100):
            yield {"input_ids": torch.tensor([i])}

config = DataLoaderConfiguration(custom_classes=(MyDataSource,))
accelerator = Accelerator(dataloader_config=config)
loader = accelerator.prepare(MyDataSource())  # returns DispatchDataLoader