Approximation beats concentration? An approximation view on inference with smooth radial kernels