On the Markov Property of Neural Algorithmic Reasoning: Analyses and Methods