Combine short segments into larger ones for better translation quality

This may be especially helpful in proverbs and psalms.
This is pealed off from https://github.com/sillsdev/silnlp/issues/250

Combine multiple short segments into 1 for training

• Assume segments are contiguous if the last segment of their id's increment by 1 and the other portion of their ID is the same.
• Combine contiguous segments into one segment first to make complete sentences and then to just extend the segment as long as the source and target lengths are < max_tokens.

Combine multiple short segments into 1 for pretranslation

• Assume contiguous segments as above
• Combine segments by the 100%, 80% rule as above, giving preference to making complete sentences
• Translate both the segments independently and combined
• Split by running Gale-church on the independently and combined pretranslations (there may be a much better algorithm based on cognates, etc. since most of the words will be exactly the same)

@ddaspit - what do you think?

The issue names no files, tests, or entry points. Start by locating the training and pretranslation pipelines, then determine how segment IDs, token limits, complete sentences, and Gale-Church splitting are represented; the work is done when contiguous short segments can be combined for both workflows and the resulting pretranslations can be split again.