Make XGBoost evaluation configurable

[amarakramali]

Summary

• make the XGBoost evaluator import-safe and reusable
• add strict feature-table, timestamp, numeric-value, and hourly-continuity validation
• configure paths, evaluation windows, and boosting rounds from the CLI
• preserve exact chronological train/validation/test partitions
• add the XGBoost evaluation to the synthetic no-download quickstart
• cover split behavior, invalid input, metrics, and plot artifacts with tests

Why

src.xgb_eval previously executed on import and used hard-coded Windows paths. The documented synthetic workflow could build features but could not run the XGBoost comparison with those sample paths, and the evaluator could not be safely reused in tests or other code.

Impact

The original defaults remain available, while users can now evaluate any compatible feature CSV and choose output locations and evaluation settings. The workflow also reports a zero-error baseline as an undefined improvement instead of failing.

Validation

• python -m pytest -q — 63 passed
• uvx --from ruff==0.16.0 ruff check src tests streamlit_app.py — passed
• python -m compileall -q src tests streamlit_app.py — passed
• full 90-day synthetic data → features → 30-day XGBoost evaluation smoke test — metrics CSV and 119 KB PNG generated
• smoke result: XGBoost MAE 289.86 MW vs baseline 337.09 MW (14.0% improvement)