Do LLMs Understand User Preferences? Evaluating LLMs On User Rating Prediction