On the Tradeoff between Energy, Precision, and Accuracy in Federated Quantized Neural Networks