Reconciling Communication Compression and Byzantine-Robustness in Distributed Learning