A Unified Framework and Dataset for Assessing Societal Bias in Vision-Language Models