On Selecting Few-Shot Examples for LLM-based Code Vulnerability Detection