avoid copying inputs when doing a `knn` or `inrange` for a matrix input

[KristofferC]

For some reason we did the conversion to static vectors "wholesale" for the whole input matrix instead of just doing it point by point.

julia> using NearestNeighbors

julia> tree = KDTree(rand(3, 100));

julia> @time input = rand(3, 10^6);
  0.036092 seconds (2 allocations: 22.888 MiB, 89.04% gc time)

# Before
julia> @time knn(tree, input, 5);
  0.368772 seconds (2.00 M allocations: 221.253 MiB, 5.09% gc time)

# After
julia> @time knn(tree, input, 5);
  0.395747 seconds (2.00 M allocations: 198.365 MiB, 17.31% gc time)

We can see we save copying that 20 mb input.