K-NN Regressor

[krstopro]

Implements k-NN regressor model and removes previous implementation of k-NN.
As usual there are few things I am not sure about:

• There is a lot of code repeated from KNNClassifier (e.g. deftransformp compute_knn). Was thinking pulling it out of both modules and putting it somewhere else, but that might make the code harder to understand.
• Labels y can be both rank-1 and rank-2. I think this is slightly different than usual, where tensors that can contain more than one feature are always rank-2. I would suggest allowing y to be only of rank 2.

Closes #255.

[krstopro]

• Labels y can be both rank-1 and rank-2. I think this is slightly different than usual, where tensors that can contain more than one feature are always rank-2. I would suggest allowing y to be only of rank 2.

I would prefer that we all agree on this before merging the pull request.

[josevalim]

I don't have an opinion. But it is always easier to relax the constraints later (i.e. allow only rank 2 for now and potentially also rank 1 later). @msluszniak do you have any thoughts?

[msluszniak]

Yes, I agree that for now we can accept only 2-dim labels

[krstopro]

Alright, will push the code with changes soon. :)

[josevalim]

Beautiful, thank you @krstopro!

Once @msluszniak updates both KNN guides we will ship a new version!!!