When to Act and When to Ask: Policy Learning With Deferral Under Hidden Confounding