Multi-Armed Bandits for Many-Task Evolutionary Optimization