Evolving Graph Learning for Out-of-Distribution Generalization in Non-stationary Environments