Add trap bundle collection/ablation utilities, tests, and bump version to 0.8.4

[charlesmartin14]

Motivation

• Provide a way to capture, persist, and reuse per-layer trap analysis artifacts (SVD, permutations, original/permute matrices) to enable deterministic ablation experiments and reproducible analysis.
• Expose high-level helpers to remove individual traps or sets of traps from models using precomputed decompositions, and to run ablation experiments across checkpoints.

Description

• Introduces a new weightwatcher/trap_bundles.py module with TrapAnalysisBundle dataclass and functions analyze_traps_bundle, save_trap_bundle, load_trap_bundle, remove_single_trap_from_bundle, remove_traps_from_bundle, and run_trap_bundle_ablation_experiment to collect, serialize, and apply bundle-based ablations.
• Extends weightwatcher/trap_analysis.py to accept an optional bundle_collector and to populate it with SVD and permutation artifacts for each analyzed layer, using svd_full and safe permute_ids handling.
• Adds tests/test_trap_bundles.py with unit tests covering round-trip bundle save/load, seed-independence of ablation vs fresh analysis, mismatch validation, and legacy API behavior; tests are skipped when torch is not available.
• Bumps package version in weightwatcher/__init__.py from 0.8.3 to 0.8.4.

Testing

• Ran pytest -q tests/test_trap_bundles.py, and tests passed on environments with torch installed and were skipped when torch was not available.
• The new bundle collection and ablation code paths exercised by the tests completed without uncaught exceptions under the test conditions.

---

Codex Task