Hard-Constrained Neural Networks with Universal Approximation Guarantees