Hallucinate Less by Thinking More: Aspect-Based Causal Abstention for Large Language Models