Learning Distributed and Fair Policies for Network Load Balancing as Markov Potential Game