Exploring Social Bias in Downstream Applications of Text-to-Image Foundation Models