Assessing Empathy in Large Language Models with Real-World Physician-Patient Interactions