Few shot clinical entity recognition in three languages: Masked language models outperform LLM prompting