Efficiently Editing Mixture-of-Experts Models with Compressed Experts