Read Like Humans: Autonomous, Bidirectional and Iterative Language Modeling for Scene Text Recognition