Prior to this change, the find function implementation for the
CoverTree class could have potentially returned the wrong result
in cases where there were multiple points in the dataset
equidistant from p. For example, the current test passed for k=3
but failed to produce the correct result for k=4 (it claimed that
3, 4, 5, and 7 were the 4 closest points to 5 in the dataset
rather than 3, 4, 5, and 6). Sorting the neighbors vector before
collecting the first k values from it resolved this issue.
