Operator-Theoretic Framework for Gradient-Free Federated Learning