SignCLIP-style "_-shot Nearest Neighbors" Evaluation

[cleong110]

Like in #21, we can have the number of samples and the ratio of in and out of class samples as a settable parameter.

We can have, for example, "10-shot, 1/5", where we take 10 from the target gloss, and 40 not from the target gloss, so the correct neighbors are 1/5 of the total. Then you run your distances, take the top k, and do KNN classificaton.

Then, report KNN accuracy!