Shortcut-connected Expert Parallelism for Accelerating Mixture-of-Experts