Add paper-aligned trap metrics, FFT/burden variants, and trap verification/remove workflow

[charlesmartin14]

Motivation

• Align trap analysis with the NeurIPS paper by producing Porter-Thomas-centered localization metrics and a clear scalar "variance burden" while preserving legacy diagnostics for compatibility.
• Provide richer experiment support (burden variant sweeps, Fourier/FFT diagnostics) to let users compare localization/burden definitions and frequency-domain behavior.
• Improve reproducibility and safety of trap removal by recording permutation identity, signing artifacts, and adding verification logic to remove_traps.

Description

• Introduces new modules: trap_burden_variants.py, trap_fourier.py, and trap_identity.py implementing burden-variant math, Fourier/FFT diagnostics, and permutation/trap identity & verification helpers.
• Extends trap analysis to return paper-facing metrics (trap_delta, trap_ipr, trap_q, trap_diffuseness, trap_top_sector_overlap, trap_variance_burden, layer_trap_variance_burden) and optional burden variants and FFT-derived columns, plus options burden_variants, return_burden_components, return_burden_raw, trap_fft and trap_fft_config.
• Adds deterministic permutation metadata (perm_signature, permutation_n, trap_seed, trap_identity_key) and verification tooling used by remove_traps; remove_traps gained parameters verify_traps, return_analyze, traps, rtol, atol, and min_vector_cosine, and can return a verification dataframe when requested.
• Refactors and consolidates trap helper functions into trap_analysis.py and hooks them into weightwatcher.WeightWatcher (new helper methods: compute_trap_delta, compute_trap_ipr_q, compute_trap_ipr_q_uniform, compute_top_sector_overlap, compute_trap_variance_burden, and compute_original_basis_for_traps), and preserves legacy heuristics for backward compatibility.
• Documentation and README updated to describe the new metrics, defaults, and recommended FFT config fields.

Testing

• Ran the trap unit tests including tests/test_analyze_traps.py (expanded to cover new paper metrics, top-sector behavior, and index conventions), and they passed.
• Added and ran tests/test_trap_burden_variants.py covering burden math and API behavior, and tests/test_trap_fft.py covering Fourier helpers and integration hooks, and these tests passed.
• Updated tests/test_remove_traps.py with verification/return_analyze tests exercising the new remove_traps behavior, and those tests passed (torch-dependent tests are skipped when PyTorch is unavailable).

---

Codex Task