Learning Stationary Markov Processes with Contrastive Adjustment