On the Universal Approximation Property and Equivalence of Stochastic Computing-based Neural Networks and Binary Neural Networks