Improving Generalizability of Extracting Social Determinants of Health Using Large Language Models through Prompt-tuning