Beyond Text-Dominance: Understanding Modality Preference of Omni-modal Large Language Models