Scalable operator allocation for multi-robot assistance: A restless bandit approach