Few-shot clinical entity recognition in English, French and Spanish: masked language models outperform generative model prompting F i n d i n g s - E M N L P