Add overfitting modes, GPU detection, and overfit-generation plotting to XGBoost notebook

[charlesmartin14]

Motivation

• Introduce deterministic overfitting scenarios so models can be systematically driven into overfit regimes for analysis and visualization.
• Prefer GPU compute when available to speed up XGBoost training while keeping a safe CPU fallback.

Description

• Added OVERFIT_MODES and apply_overfit_mode to provide named overfit perturbations to base XGBoost parameters (deep trees, high learning rate, no regularization, full sampling, long training).
• Implemented detect_xgb_compute_params() which probes XGBoost to choose GPU vs CPU XGB_COMPUTE_PARAMS and exposes XGB_COMPUTE_BACKEND with a printed status.
• Extended fit_and_score() to accept case_type, overfit_mode, and seed_offset; use a derived run_seed for data splits and model seeds, apply overfit parameter modifications and fixed round mappings for overfit runs, and add accuracy_gap, case_type, and overfit_mode to results.
• Updated WeightWatcher invocation to use run_seed and added a plotting cell that generates overfit runs from base (good) models and plots histograms for alpha, ERG_gap, and num_traps using OVERFIT_REPEATS_PER_MODEL and PLOT_METRICS.

Testing

• Executed the GPU detection probe which completed and reported a detected backend successfully.
• Performed a smoke fit_and_score run on a small dataset end-to-end (train → convert → WeightWatcher analyze) which completed without error.
• Observed that long-running cross-validation on full datasets can be interrupted (previous xgb.cv runs triggered a KeyboardInterrupt), indicating large CV jobs may still be time-consuming in this environment.

---

Codex Task