Unit test human evaluation

[Stanislaw Jastrzebski]

Add unit test, just to err on the safe side

[Stanislaw Jastrzebski]

This should test both the distance calculation, ordering, bucketing, and mixing. All parts independently and jointly.

[Stanislaw Jastrzebski]

I am not sure if compute_distances and compute_distances_fast are actually equivalent. There is some issue with buckets having incorrect triplets (e.g. close ones)? Extremely weird and worrying.

Idea: compute_distances_fast with batch_size=1 and batch_size=100.

[Stanislaw Jastrzebski]

#77