Adds two more examples of croped word recognition, and show how to set the beam size parameter for small lexicon.

OCRBeamSearchDecoder code refactor: fixes most of the TODOs comments, cleans-up code, and fix a bug in score_segmentation(...) member function.

Add benchmark code for the Chars74k dataset. Using the CNN character classifier reaches 75% and 84% accuracy for case-sensitive and case-insensitive recognition respectively.