Better saving of chrf3 for quality estimation

[mshannon-sil]

We should save the projected chrf3 scores alongside the usability scores in the verse, chapter, and book usability files. Currently only the verse scores are being saved, and in their own separate file which makes it more difficult to see them side by side with usability.

As part of this, we should consider the best way to obtain projected chrf3 at a chapter and book level. Currently, the verse confidences are used to get a projected chrf3 and then usability at the verse level. And then the verse level usabilities are averaged to get chapter and book usabilities. However, we already have chapter and book confidences being saved during the test step (calculated as the geometric mean of the verse confidences), so a possibly better alternative would be to project chrf3 and usability for those chapter and book confidences directly. I'd like to do some comparison to see how different the resulting usabilities would be using this method compared to how it works right now.

[mshannon-sil]

The results are pretty similar, just slightly different (~1pt), when projecting usability directly from chapter/book level confidence and their projected chrf3 scores, compared to our current method of averaging the usability scores of the verses that make up the chapter/book. However, I think since we're trying to ask for the usability of a chapter/book based on how usable each verse is, and also since the chrf3/usability distribution has only been modeled with examples of verses, I think it's best to keep it how we've had it and continue to calculate usability as the average of the usability of the verses.

[benjaminking]

I'd guess that the difference is simply due to the difference between the geometric and arithmetic mean. But I agree with keeping the calculation as is, especially since usability can be interpreted as a proportion of usable verses.