Field Notes on Scaling MoE Expert Parallelism with DeepEP