Mixture-of-Transformers: Sparse and Scalable Architecture for Multi-Modal Models