Deceive, Detect, and Disclose: Large Language Models Play Mini-Mafia