Verify when Uncertain: Beyond Self-Consistency in Black Box Hallucination Detection