Implement universal _find_neighbors function using ballTree and KDTree

[paco-barreras]

There is an attempt at implementing this already

• Find and push branch with BallTree implementation for dbscan
• Remove redundant hdbscan _find_neighbors and test
• Optionally return KDTrees
• Consolidate into a universal function for dbscan, seqscan, and hdbscan with optional returns

[paco-barreras]

BallTree and KDTree preliminary implementations are in branch core_distances_with_ball_tree.

These are faster than the previous implementation based on a dense pairwise comparison. It seems to me that

preprocessing._find_temp_neighbors
preprocessing._find_neighbors

differ only on whether we want to incorporate spatial thresholding or not. While hdbscan._find_temp_neighbors is simply redundant. We could have a single function: _find_neighbors, with time_only = False as a boolean flag for whether we want time only. It will also be convenient to return the output as a list of edges (time_pairs seems to be the current name), or a dictionary, optionally. However, it would seem that a natural and optimized way to store the output is a networkx object.

We might also want to just output the kdtree objects directly, what possible advantage could we get from instantiating the graph? perhaps to reuse distance calculations? It also seems quite redundant.

@andresmondragont , did you already have some implementation of this? trying to unify everything?

[paco-barreras]

Fixed in #245