On Fairness of Unified Multimodal Large Language Model for Image Generation