Add randomized correlation-trap analysis (analyze_traps) with trap metrics and heuristics

[charlesmartin14]

Motivation

• Provide a reproducible, randomized workflow to detect and inspect correlation "traps" (isolated spike modes) in layer weight matrices.
• Expose per-trap vector metrics and order-invariant statistics to characterize localization vs diffuseness of traps.
• Add a compact heuristic to assess trap severity (diffuseness / risk) and tooling to visualize traps for debugging and model surgery.

Description

• Introduces WeightWatcher.analyze_traps(...) and supporting methods including apply_analyze_traps, analyze_single_trap, identify_trap_mode_indices, compute_original_basis_for_traps, apply_trap_mp_fit, and _trap_result_columns to return one DataFrame row per detected trap.
• Adds per-trap vector metrics (_trap_vector_metrics) and order-invariant statistics (_trap_vector_order_invariant_stats) plus helpers (_gini_abs, _topk_mass_fractions) and per-trap aggregation into the result rows.
• Implements assess_trap_diffuseness(...) to compute trap_diffuseness_score, trap_risk_score, and trap_assessment, and adds plotting via plot_trap_analysis.
• Makes permutation deterministic/reproducible by adding an optional rng parameter to permute_matrix(...) and threading rng through apply_permute_W and analyze_traps.
• Adds unit tests: tests/test_analyze_traps.py (exercise API, metrics, reproducibility, columns) and tests/test_trap_diffuseness.py (heuristic behavior).
• Updates documentation and packaging: bumps __version__ to 0.8.0, updates README.md with usage examples for analyze_traps, and records new changes in Changelog.txt.

Testing

• Ran the new unit tests tests/test_analyze_traps.py and tests/test_trap_diffuseness.py via the repository test runner (python -m unittest / pytest); the tests executed and passed in the development environment.
• Smoke-tested trap plotting/save behavior via analyze_traps(..., plot=True, savefig=...) to ensure no exceptions when saving figures.

---

Codex Task