Manifold learning notebooks

[msluszniak]

I decided to replace NNDescent in Trimap with LargeVis. In the original implementation of Trimap, authors used Annoy, which is more similar to LargeVis and it's faster. In Google's implementation using Jax, they use NNDescent, which is slower and eventually makes the whole implementation slower than the not-optimised, original one.
original implementation: https://github.com/eamid/trimap
Google implementation: https://github.com/google-research/google-research/tree/master/trimap

[msluszniak]

@krstopro do you know there is any way to add support in LargeVis or precisely in RPF for data that is in test cases here and currently fails?

[krstopro]

@msluszniak Yeah, I should fix that. The main reason I didn't do it in the first place was because I wouldn't recommend using random projection forest for those cases, i.e. when the sample size is less than two times leaf size.

I would recommend making k-NN algorithm configurable for Trimap. Similarly as it was done in KNNClassifier and KNNRegressor.

[msluszniak]

Yeah, that's a good idea

[josevalim]

I would recommend making k-NN algorithm configurable for Trimap. Similarly as it was done in KNNClassifier and KNNRegressor.

Good idea indeed, I have created an issue for it here: #280.

#272 is also related here (it could be another example to show in the notebook).

[krstopro]

Looks great, few comments on the writing from my side.

[josevalim]

Ship it once CI passes!