Add multi-OCR mode

[wrznr]

As things stand, multiple hypotheses for (lines of) characters will only be available through multiple runs of OCR in OCR-D. What is missing is a clever way to serialize a single text version out of them. cor-asv-ann has the means available to provide this disambiguation step.

[bertsky]

This is definitely an objective. I have long-standing plans to integrate the ideas of http://www.ccs.neu.edu/home/dongrui/ocr.html. But this may need some time to harmonize, as the basic NMT implementations are quite different, and this would add a completely complementary training/data regime.