CoMPosT: Characterizing and Evaluating Caricature in LLM Simulations