Use PyMarian for COMET evaluations

[marco-c]

It is supposedly 8x faster than what we're using according to https://arxiv.org/pdf/2408.11853.

[gregtatum]

This is probably a bit lower priority than other cost & perf issues as the evaluation tasks don't take too long to run, for instance on the hr-en run it was just 10 GPU machine hours compared to ~5 days each for training the student and teacher.