A Universal Approximation Theorem of Deep Neural Networks for Expressing Probability Distributions