Double Quantization for Communication-Efficient Distributed Optimization