Refactor trap workflow API: add randomize_model and propagate permuted_ids

[charlesmartin14]

Motivation

• Decouple the trap workflow so permutation/randomization can be performed independently and reused to avoid re-randomization, sync issues, and repeated work across analysis/removal stages.
• Allow downstream trap analysis and removal to accept a pre-randomized model and deterministic permuted_ids to guarantee consistent trap indexing and fingerprints across tools.

Description

• Added randomize_model(ww, model, ...) in weightwatcher/remove_traps.py which deep-copies and permutes eligible layer weights and returns (randomized_model, permuted_ids).
• Exposed a public WeightWatcher.randomize_model(...) method that forwards to the new helper for callers to obtain a randomized model and per-layer permutation ids.
• Extended WeightWatcher.analyze_traps(...) signature to accept randomized_model and permuted_ids, and updated trap_analysis.analyze_traps(...) to accept and forward permuted_ids into per-layer params.
• Updated apply_analyze_traps to use externally supplied permuted_ids when present instead of generating new permutations so analysis and removal remain synchronized.
• Files changed: weightwatcher/remove_traps.py, weightwatcher/weightwatcher.py, and weightwatcher/trap_analysis.py.

Testing

• Ran targeted unit tests: pytest -q tests/test_analyze_traps.py::TestAnalyzeTraps::test_analyze_traps_trap_burden_columns_appear tests/test_remove_traps.py::test_remove_traps_public_api_direct_call which produced 1 passed, 1 skipped.
• No other automated tests were run in this change set.

---

Codex Task