Add sanity-check notebook for analyze_traps with injected localized singular mode

[charlesmartin14]

Motivation

• Provide a reproducible, minimal experiment to validate analyze_traps detection of a single injected localized singular mode in a dense matrix.
• Capture expected localization and top-k mass metrics for a controlled left/right singular vector pair to catch regressions in trap detection.
• Include a runnable example that demonstrates plotting and artifact return paths of the analyze_traps workflow.

Description

• Add notebooks/analyze_traps_iid_gaussian_sanity.ipynb which builds a 200 x 400 IID Gaussian matrix and injects one highly localized singular vector pair and an amplified singular value.
• Wrap the injected matrix in a simple OneLayerMatrix nn.Module, then call watcher.randomize_model and watcher.analyze_traps to detect traps and compute left_top_mass, right_top_mass, and top_5_mass metrics.
• Compute expected metrics (top_k_mass, localization_ratio) for the injected vectors, assert close agreement with observed trap metrics, and exercise analyze_traps(plot=True) to generate plots.

Testing

• Executed the notebook end-to-end; the embedded assertions that compare observed vs expected top-5 mass and localization passed successfully.
• Ran analyze_traps with return_artifacts=True and plot=True in the notebook to ensure both code paths execute without error.
• No unit test suite changes were made outside the notebook and no CI test failures were introduced.

---

Codex Task