Revisiting Hallucination Detection with Effective Rank-based Uncertainty