Evaluating Large Language Models as Generative User Simulators for Conversational Recommendation