Cognitive bias in clinical large language models