End-to-End Scene Text Spotting at Character Level