SAC$^3$: Reliable Hallucination Detection in Black-Box Language Models via Semantic-aware Cross-check Consistency