Separator Injection Attack: Uncovering Dialogue Biases in Large Language Models Caused by Role Separators