Additive Noise Annealing and Approximation Properties of Quantized Neural Networks