Tighter Prediction Intervals for Causal Outcomes Under Hidden Confounding