Add bulk-mode detection and removal to remove_traps (explicit and random bulk vectors)

[charlesmartin14]

Motivation

• Extend the trap-removal workflow to also identify and remove bulk rank-1 vectors, and to support removing either explicit bulk vectors or a requested number of randomly-chosen bulk vectors.

Description

• Add identify_bulk_mode_indices to classify bulk modes (<= TW threshold) in a layer using the same MP/TW logic as traps.
• Add collect_bulk_artifacts to extract bulk-mode artifacts via the existing permute + MP/TW analysis pipeline.
• Extend apply_remove_traps and remove_traps to accept new parameters bulk_indices and num_random_bulk_vectors, to validate indices, and to support removing both traps and bulk vectors by replacing their rank-1 contributions with make_stat_matched_random_matrix draws.
• Update WeightWatcher wrappers (apply_remove_traps and remove_traps) to expose the new bulk_indices and num_random_bulk_vectors arguments and update docstrings accordingly.

Testing

• Ran the new unit tests tests/test_remove_traps.py::test_remove_traps_can_remove_explicit_bulk_vector and tests/test_remove_traps.py::test_remove_traps_can_remove_random_bulk_vector, and the existing test_remove_traps_invalid_indices_warns_and_skips, and they all passed under pytest.

---

Codex Task