Text this: Evaluating large language models for multilingual vulnerability detection at dual granularities