Learning from Binary Multiway Data: Probabilistic Tensor Decomposition and its Statistical Optimality