An Efficient Gradient-Aware Error-Bounded Lossy Compressor for Federated Learning