Fix issue where precomputed knn aren't being reused

[cannoneyed]

Fixes #2082.

[wchargin]

@manivaradarajan: Let’s make sure to get this in for the 1.14 release.

[wchargin]

Thanks. This looks correct, and appears to fix the divergence for me.
Before this commit, running t-SNE with perplexity=8 then re-running with
perplexity=100 and waiting for a minute or so caused clusters to spiral
farther and farther apart. After this commit, it behaves correctly.

Almost a shame—the wrong behavior was actually really pretty. :-)

[cannoneyed]

@wchargin I wound up fixing another knn-related bug in this PR (that @nsthorat found while debugging the standalone release) - You'll probably want to give it another look.

tl;dr is there was some slight drift between the umap.d.ts file and the umap implementation's function signature that was causing an issue with small dataset sizes.

[wchargin]

Ah, didn’t see your comment before starting the review. Thanks for the
context. Please separate the other fix into another PR.

[cannoneyed]

OK cool, I'll move the other fix into a separate PR

[cannoneyed]

New PR opened here

[wchargin]

Great; thank you! Ideally the whitespace changes on otherwise-unchanged
lines in this PR should be omitted, too—what I meant by my comment was
“please avoid introducing new whitespace errors”. If we could keep
future PRs localized, that’d be great.