Add BLEURT metric

[patrick-wilken]

BLEURT is used as a metric at the IWSLT subtitling track. This is not the code used for evaluation there, but it probably makes sense to add this implementation here for convenience. Scores should be similar to the official ones if one uses --metrics AS-BLEURT and a reference in SRT format. This will Levenshtein-align the hypothesis words into the reference subtitle blocks and then compute and average BLEURT scores over the aligned pairs. For IWSLT, mwerSegmenter is used instead, resulting in slightly different alignment compared to our Levenshtein implementation.
To me it is still an open question whether calculating learned metrics on such automatically aligned segments is really sound, because in many cases starts and ends of sentences are chopped off. For surface-level metrics that does not matter much, but learned metrics are trained on proper complete sentences. And indeed recent work by Post and Hoang shows that - in this case - COMET scores degrade with automatic segmentation, although system ranking seems to be preserved overall.
I would also intuitively expect that full sentences as reference segments rather than subtitle blocks should give more accurate scores. (Can be done by passing a plain text reference, see README.) But as of now I did not evaluate that systematically.
So, even if we merge this, the BLEURT metric might be subject to changes/improvements - as opposed to the other metrics which I want to be very stable.

Note, that many other learned metrics like COMET require a source input and are therefore not trivial to compute for subtitling, or speech translation in general.