Integrating Physician Diagnostic Logic into Large Language Models: Preference Learning from Process Feedback