Initialize indices with 1, allows for Infs and NaNs in the data

[nlw0]

The patch merely modifies the algorithm to initialize the index list with 1 instead of -1. New tests are added that reproduce issues #78 and #125.

It's a bit of a decision what to do with NaNs, in my opinion it's best if we ensure indices are always valid, and a NaN distance is as good as Inf.

[nlw0]

Tried out NaN on data, and turns out in that case we still have problems. Hard to figure out what to do in that case, will need more work, not as easy as this patch. Left a commented test for that issue.

[nlw0]

Turns out this patch actually does not solve #78, we return a valid index, but it's not the correct one in that case... Not sure how to solve it. for #125 it's fine.

[axsk]

Not sure why you commented this out,
but if its merely because the test fails, a @test_broken might be appropriate so we know its broken :)

[axsk]

I guess this might also solve #127 (maybe add the test)

Considering #78 this might be related to something else (since it only happens with BallTrees) so I think it would be nice to have this merged either way.