Auto-split batches for retranslation

[machinewrapped]

I've found that GPT is much more likely to desync with larger batches, and splitting the batch in two and retranslating almost always fixes the problem. It would make sense from a usability and economic perspective to do this automatically when batch validation fails.

Splitting the lines in two at the longest gap should be a reasonable approach (perhaps weighted according to proximity to the midpoint - splitting off one or two lines probably won't help).