Estimating Knowledge in Large Language Models Without Generating a Single Token