Compression with Exact Error Distribution for Federated Learning