fix: resolve pruning pipeline bugs

[nh13]

Summary

Fixes several bugs in the dataset pruning pipeline (prune_dataset.py), plus an incorrect CLI description in edit_dataset.py.

• Vestigial fold loop — calculate_pruning_thresholds wrapped its entire body in for fold in range(NUM_FOLDS) but unconditionally returned on the first iteration, so the loop was dead code (and misleading). Removed it.
• Wrong loader argument — make_data_loader was passed training_params.num_epochs where num_workers is expected, which could launch hundreds of worker processes.
• Operator precedence — the cyclic validation-fold index was pruning_fold + 1 % len(fold_datasets), which parses as pruning_fold + (1 % len) and indexes out of range on the final fold. Fixed to (pruning_fold + 1) % len(fold_datasets).
• Division-by-zero guards — added informative ValueErrors when no data pass the confidence thresholds, and when the summed error rates reach 1.0 (which would otherwise divide by zero and violate the rank-pruning assumptions).
• Return annotation — generated_pruned_data_for_fold is a generator; corrected List[int] → Generator[Datum, None, None].
• Wrong CLI descriptions — both prune_dataset.py and edit_dataset.py had argparse descriptions reading "train the Mutect3 artifact model".

Test plan

• ruff format / ruff check clean; both modules compile and import.
• Note: the test_prune_dataset integration test requires an EXPERIMENTAL_MODEL checkpoint that is not committed to the repo, so it cannot run in CI. The pruning logic was verified by inspection and targeted checks (generator semantics, cyclic fold indexing, guard conditions).