Metalearning Linear Bandits by Prior Update