Add correlation-trap analysis (analyze_traps) with heuristics, plotting, and tests

[charlesmartin14]

Motivation

• Provide a dedicated workflow to detect and inspect randomized/permuted "correlation traps" in layer weight matrices using MP-fit and rank-1 isolation.
• Expose a reproducible API to analyze traps layer-by-layer, evaluate trap vectors in original basis, and classify trap severity with a simple heuristic.

Description

• Implemented analyze_traps on WeightWatcher along with helpers: apply_analyze_traps, apply_trap_mp_fit, identify_trap_mode_indices, compute_original_basis_for_traps, analyze_single_trap, assess_trap_diffuseness, plot_trap_analysis, _trap_vector_metrics, and _trap_vector_order_invariant_stats to perform permutation, MP bulk-fitting, trap isolation, unpermutation, metric computation, and plotting.
• Added trap-specific metrics to the returned row dictionary (e.g., sigma_perm, eval_perm, mp_bulk_max, various u_/v_ metrics, trap_diffuseness_score, trap_risk_score, and trap_assessment) and ensured details ordering is consistent.
• Added README example usage for trap analysis showing watcher.analyze_traps(...) and image saving.
• Added unit tests: tests/test_analyze_traps.py to validate API, dataframe columns, reproducibility and vector/order-invariant statistics, and tests/test_trap_diffuseness.py to validate assess_trap_diffuseness behavior and outputs.

Testing

• Added tests/test_analyze_traps.py and tests/test_trap_diffuseness.py and executed them with pytest -q against the modified codebase; tests passed where dependencies were available.
• Torch-dependent tests in tests/test_analyze_traps.py are guarded with @unittest.skipUnless(TORCH_AVAILABLE, ...) and will be skipped automatically when torch is not installed.
• The new unit tests exercise metric computation, ordering, reproducibility with fixed seeds, and the diffuseness/risk heuristics and reported results during the run.

---

Codex Task