Caesar: A Low-deviation Compression Approach for Efficient Federated Learning