L-GreCo: Layerwise-Adaptive Gradient Compression for Efficient and Accurate Deep Learning