Uncertainty-Aware Instance Reweighting for Off-Policy Learning