Add custom iterable dataloader support in Accelerator.prepare

[t7r0n]

Summary\n- add DataLoaderConfiguration.custom_classes so users can register iterable dataloader-like classes\n- add a lightweight CustomIterableDataLoader wrapper that:\n - iterates the underlying object\n - optionally moves each batch to device\n - preserves accelerate dataloader state tracking for gradient synchronization\n- update Accelerator.prepare / prepare_data_loader flow to treat configured custom classes as dataloaders\n\n## Tests\n- add test_prepare_data_loader_with_custom_iterable_loader in tests/test_data_loader.py\n- add test_prepare_with_custom_iterable_dataloader in tests/test_accelerator.py\n\n## Local verification\n- python3 -m compileall src/accelerate/accelerator.py src/accelerate/data_loader.py src/accelerate/utils/dataclasses.py tests/test_data_loader.py tests/test_accelerator.py\n- direct runtime smoke check with PYTHONPATH=src for both prepare_data_loader and Accelerator.prepare custom iterable paths\n- full pytest run in this environment is blocked by missing test dependency: parameterized\n\nCloses #2975

[t7r0n]

Additional local validation run in a uv-managed virtual environment: pytest tests/test_data_loader.py tests/test_accelerator.py -k not setting_cpu_affinity. Result: 81 passed, 5 skipped, 1 deselected. The deselected CPU affinity test is hardware-specific on this machine (NVML CPU affinity not supported) and unrelated to this PR changes.