Beyond Overcorrection: Evaluating Diversity in T2I Models with DivBench