fix: resolve pruning pipeline bugs

[nh13]

Summary

• num_epochs passed where num_workers expected in make_data_loader, potentially launching hundreds of worker processes
• calculate_pruning_thresholds had a vestigial for fold in range(NUM_FOLDS) loop that always returned on the first iteration; removed the loop
• Added division-by-zero guards with informative ValueError messages in error rate calculations
• Fixed operator precedence bug: a + b % c evaluated as a + (b % c) instead of (a + b) % c for fold index
• Fixed return type annotation: List[int] -> Generator[Datum, None, None]
• Fixed wrong argparse descriptions in prune_dataset.py and edit_dataset.py (both incorrectly said "train the Mutect3 artifact model")

Test plan

• test_prune_dataset integration test passes
• Verify pruning runs to completion on example data