Optimal Treatment Regimes for Proximal Causal Learning