Foundations of Deep Learning: Compositional Sparsity of Computable Functions