Communication-Efficient Federated Learning by Quantized Variance Reduction for Heterogeneous Wireless Edge Networks