Fix multi-gpu inference

[arturnn]

This PR fixes multi-gpu inference, as reported in #177
While trying to fix it, I found some more errors related to proper handling of error spans metadata (with XCOMET models).

I tried out the fixes with XCOMET-XL and wmt22-comet-da models - but it could use some more testing (the code could also be somewhat improved, but I haven't mastered the COMET codebase yet ;) )

[ricardorei]

Hi @arturnn sorry for taking time to do this I was out in December.

I just merged your code and published a new version (2.2.1). I tested you code with both Unbabel/wmt22-comet-da and Unbabel/XCOMET-XL. On both models everything seems to be working well.

I used the following files:
src.txt
ref.txt
mt.txt

The output and scores using 1 GPU or 2 GPUs is exactly the same and thus everything seems correct:

comet-score -s src.txt -t mt.txt -r ref.txt --model Unbabel/XCOMET-XL
comet-score -s src.txt -t mt.txt -r ref.txt --model Unbabel/XCOMET-XL --gpus 2

Output:
/mnt/data/ricardorei/external/tmp/mt.txt score: 0.8700