Geospatial Mechanistic Interpretability of Large Language Models