Expose DataLoader kwargs (TrainSettingsConfig.dataloader_kwargs) + add validation helper reject_reserved_keys - #3554

…tings

Training on small datasets over hundreds of epochs respawns DataLoader
workers every epoch (persistent_workers=False) with a fixed prefetch depth
of 2, leaving GPU utilization on the table for CPU-heavy pipelines. These
knobs existed in PyTorch but were not reachable through TrainSettingsConfig,
pytorch_config.yaml, or pytorch_cfg_updates (strict schema rejects them).

- TrainSettingsConfig: add dataloader_persistent_workers (default False)
  and dataloader_prefetch_factor (default None = PyTorch default 2,
  validated >= 1).
- train(): forward both to the training DataLoader when
  dataloader_workers > 0; warn and ignore them for single-process loading
  (PyTorch requires workers for both options). .get() with defaults keeps
  old persisted configs without the keys working.
- base.yaml / base_detector.yaml defaults + pytorch_config.md docs.

Adds config roundtrip/validation tests and train() DataLoader-kwarg tests.

Addresses #3547