On Dimension-Independent Approximation by Neural Networks and Linear Approximators