Efficient Deviation Types and Learning for Hindsight Rationality in Extensive-Form Games