Add full Covertype hyperparameter sweep notebook with Colab + WeightWatcher integration

[charlesmartin14]

Motivation

• Provide a runnable Colab-ready notebook to sweep/tune XGBoost on the UCI Covertype dataset and export WeightWatcher diagnostics for representative models.
• Make the sweep robust to different runtimes and large stacked multiclass matrices and add graceful fallbacks when xgboost2ww/WeightWatcher conversion fails.

Description

• Replace and extend notebooks/XGBoost2WWCovertypeHyperparameterSweep.ipynb with a complete Colab-friendly notebook that includes a Colab badge, Drive-mount logic, optional pip install cell, and notebook-level runtime knobs (USE_SAMPLING, SAMPLE_N, SWEEP_SEEDS, T_POINTS_SWEEP, WW_MAX_N, etc.).
• Add deterministic data loading with a fetch_openml fallback, stratified train/val/test split helpers (build_split) and reproducible training/evaluation functions (fit_eval_config_once, fit_eval_config) that return accuracy/log-loss/F1 plus WeightWatcher metrics.
• Integrate xgboost2ww.convert + WeightWatcher analysis with robust try/fallback handling (fallback to multiclass='avg' and reduced eval settings), include convert diagnostics and torch output handling, and increase WW matrix cap (WW_MAX_N).
• Add a debug cell that trains one config, prints training/test stats, converts to torch, and runs deep WeightWatcher diagnostics (power-law fit, layer info and plots); many example outputs from a debug run were captured in the notebook.

Testing

• Executed the initial Colab mounting and checkpoint setup cells successfully (Drive mounted and checkpoint paths printed).
• Ran the debug training cell: model trained with early-stopping and printed best_iteration, train_acc and test_acc, and xgboost2ww.convert returned a torch Sequential with expected weight shapes.
• Ran WeightWatcher analysis on the debug model and observed successful analysis logs (power-law fit started) and captured metrics/plots without crashing; the notebook shows these diagnostics as outputs.

---

Codex Task