Benchmarking and Adapting On-Device Large Language Models for Clinical Decision Support