Predicting emergency department disposition using machine learning and large language models to support proactive capacity management: a multicenter retrospective study