Refactor notebooks to run analysis in Python cells and add labeled plotting helpers

[charlesmartin14]

Motivation

• Ensure notebooks do not execute shell/terminal commands and instead load experiment artifacts and run analyses within Python cells for reproducibility and safety.
• Provide reusable helpers to compute summary statistics and produce labeled plots with 95% CI error bars so notebooks show well-labeled, replicate-aware visuals.

Description

• Added reusable analysis helpers to src/wwgpt/analysis.py: errorbar_indices, mean_ci95, collect_metrics, and plot_metric_curve, which normalize metrics, compute CI/error-bar indices, and produce labeled optimizer plots with 95% CI error bars.
• Replaced a minimal unlabeled validation-loss plot in analyze_results with the new plot_metric_curve helper to produce labeled plots with confidence intervals.
• Refactored notebooks (notebooks/01_validate_repository.ipynb, 03_weightwatcher_analysis.ipynb, 04_scaling_laws.ipynb, 05_overfitting_and_generalization.ipynb, 06_summary_report.ipynb) to load data and run analysis entirely inside Python cells, to use the shared helpers, and to avoid shell escapes or subprocess calls.
• Added notebook metadata normalization (cell id, outputs, and execution_count) so notebooks validate and compile cleanly.

Testing

• Scanned all notebook code cells for shell-escape patterns and confirmed no shell-invocation patterns were present with the custom scan script (passed).
• Compiled and validated every notebook code cell by compiling each cell source (all code cells compile successfully).
• Ran the test suite with pytest -q, which passed: 19 passed (one notebook-related warning was addressed by normalizing notebook metadata).

---

Codex Task