Communication-Efficient Distributed Learning via Lazily Aggregated Quantized Gradients