ONCE: Boosting Content-based Recommendation with Both Open- and Closed-source Large Language Models