FedDQ: Communication-Efficient Federated Learning with Descending Quantization