Quantum learning of classical stochastic processes : The completely-positive realization problem