[python][torch] Support distributed iterable dataset sharding

[XiaoHongbo-Hope]

What changed

• Shard streaming Torch datasets across DDP ranks and DataLoader workers.
• Resolve the DDP context at iteration time from torch.distributed, then torchrun environment variables.
• Reject a binding read limit with multiple ranks to avoid uneven DDP input.
• Make shuffled reads reproducible per epoch, rank, and worker.
• Allow rank detection to be disabled with auto_detect_rank=False.

Reference

The auto_detect_rank API and its default follow Lance Torch LanceDataset.

Validation

• 41 Torch tests passed.
• CPU torchrun test passed with 2 ranks and 2 DataLoader workers, with complete and non-overlapping split assignment.
• flake8 and git diff --check passed.