Add `similartiy` utility.

[koaning]

Add `similartiy` utility.

[koaning]

Then the Prodigy recipe might use something like:

from prodigy.sorters import ExpMovingAverage, prefer_low_scores

def make_scored_stream(stream, anchors):
    for batch in batched(stream):
        batch_text = [b['text'] for b in batch]
        distances = calc_distance(batch_text, anchors, pipeline)
        for score, ex in zip(distances, batch):
            yield score, ex

def sorted_stream(stream):
    return prefer_low_scores(ExpMovingAverage(stream))

Worth rethinking though. Something about recalculating the anchors feels a bit wasteful.