From Noisy Fixed-Point Iterations to Private ADMM for Centralized and Federated Learning