Evaluating Fairness in Large Vision-Language Models Across Diverse Demographic Attributes and Prompts