fix(train): disable persistent workers when no workers are used

[Excelius-Wang]

Since #10036 enabled persistent DataLoader workers by default, setting --dataloader_num_workers 0 raises ValueError: persistent_workers option needs num_workers > 0 when the training loader is created. This also affects the Windows default, which selects zero workers.

Disable persistence during training-argument initialization when the resolved worker count is zero. Positive worker counts retain the new default and the existing explicit opt-out. Update the argument documentation in both languages to describe the effective default.

Validation:

• Six CPU unit tests cover real HF map/iterable datasets, training/eval/prediction loaders, both training-argument classes, simulated Windows defaults, explicit persistence settings, and existing invalid-prefetch validation. Nine subcases fail on unpatched main and pass with the fix.
• Ran the actual SwiftSft pipeline with a local tiny Qwen2 model on public SQuAD, CMRC2018 and HumanEval samples, using both regular and streaming loading. All six zero-worker configurations reproduce the startup failure on main and complete training/evaluation with the fix. Losses and final model weights exactly match main with persistence explicitly disabled.
• Compared main and the patch with 1/2 workers and persistence enabled/disabled; losses and final weights match. Pre-commit passed for all four changed files.

Testing used Linux CPU and randomly initialized model weights; native Windows, GPU/distributed training and full RLHF runs were not tested.