Measuring Psychological Depth in Language Models