Zonotopes and Kalman observers: Gain optimality under distinct uncertainty paradigms and robust convergence