Performance of five free large language models in dental trauma: a 30-day longitudinal benchmark study