Can Large Language Models Explain Flight Safety Events? A Prior-Guided Semantic LLM-based Approach