Federated Learning in Open- and Closed-Loop EMG Decoding: A Privacy and Performance Perspective