Adapting Large Language Models to Log Analysis with Interpretable Domain Knowledge