Inference on Optimal Policy Values and Other Irregular Functionals via Smoothing