An Information-Theoretic Analysis of Nonstationary Bandit Learning