Non-Stationary Bandit Learning via Predictive Sampling