This AI Paper from ByteDance Introduces MegaScale-Infer: A Disaggregated Expert Parallelism System for Efficient and Scalable MoE-Based LLM Serving