WIP: Allow more types in the tree data

[KristofferC]

This is a work in progress to resolve #13 and an alternative to #14

TODO:

• Check for any performance regressions. Since this mess around with the type system we have to make sure that no type instability got introduced.
• Add test cases, maybe @tlnagy can help with that.

Creating the tree works now at least:

julia> data = rand(1:4, (10, 10^4));

julia> balltree = BallTree(data, Hamming())
NearestNeighbors.BallTree{Int64,Float64,Distances.Hamming}
  Number of points: 10000
  Dimensions: 10
  Metric: Distances.Hamming()
  Reordered: true

cc @tlnagy

[tlnagy]

Awesome. Thanks for working on this. I'll brainstorm on a couple different test cases. Should I open a new pull request for the allow_int branch?

[tlnagy]

I would like to implement at least the following tests:

• compare results on a randomly generated dataset to the distances reported by Distances.pairwise for a subset of points
• compare results to a fixed dataset without an unambigious nearest neighbor order (i.e. multiple neighbors of the same distances away)

[KristofferC]

Your plan for adding tests sounds good.

I haven't done much testing with this branch on actually running any range searches with the integer tree. I just tested that I could create the tree and that I didn't break any previous functionality.

If you get the tests to work then that's great and you can open a PR against this branch. If you get problems even running basic searches then post here with a repo and I will fix it.

[KristofferC]

Bump, still interested in this @tlnagy?

[tlnagy]

I still am. Sorry, but I've been really busy recently. Will try and look into this soon.

[KristofferC]

Superseeded by #29