Publish word with their confidence from the predictions

Depends #42 (closed)

Expose a new configuration parameter --words to publish new words based on their length under the lines. This will trigger print_word_confidence_scores as a decoding argument. The predicted output is different though.

[conf1, conf2, conf3] word1 word2 word3

We need to parse this accordingly and:

• reconstruct line transcription and publish it with the right confidence score (weighted average of its words, based on their length)
• split the line's polygon based on the words' length, same height but different widths
• call create_element_transcriptions to publish them as children of each line