Benchmarking Large Language Models on Communicative Medical Coaching: A Dataset and a Novel System F i n d i n g s - A C L