Fix duplicate samples in multi-worker streaming dataloaders

[Travor278]

This PR fixes the worker-sharding bug from #200.

When VQADataset.iter_for_worker() was used with a streaming dataset and DataLoader(num_workers > 1), each worker would walk the full stream independently. That meant every sample was processed once per worker instead of once overall.

What changed:

• shard streaming datasets per worker via .shard(...) when the underlying dataset supports it
• fall back to stride-based sharding for iterable datasets that do not implement .shard()
• add a real multi-worker regression test that reproduces the old duplication and verifies the fixed behavior
• add a small benchmark and a training-path smoke test to compare duplicate counts and effective throughput before vs after the fix

Why this matters:

• it removes duplicated samples during multi-worker streaming training
• it makes effective throughput line up with actual useful sample coverage
• it keeps the fix robust even when the iterable dataset is not a Hugging Face streaming object

Validation:

• D:\Dev\conda-envs\py313\python.exe -m unittest tests.test_worker_sharding -v
• D:\Dev\conda-envs\py313\python.exe tests\benchmark_worker_sharding.py --workers 4 --samples 2000 --repeats 5
  • before: 40000 processed / 10000 unique / 30000 duplicates / 554.33 effective samples/s
  • after: 10000 processed / 10000 unique / 0 duplicates / 900.49 effective samples/s
• D:\Dev\conda-envs\py313\python.exe tests\smoke_worker_sharding_train_path.py --device auto
  • legacy path: 288 total / 96 unique / 192 duplicates / 3.00x
  • fixed path: 96 total / 96 unique / 0 duplicates / 1.00x

Fixes #200.