Why does co_occurence not use a pre-calculated spatial graph #1165

Description of feature

As far as I can tell gr.co_occurrence uses a brute-force appraoch to find pairs of cells within a certain distance.

Couldn't that be implemented much more efficiently when relying on a pre-computed neighborhood graph? Of course, co-occurence would be limited to whatever is the max radius that was set when calculating the graph.

That way, multiple samples could also easily be respected as library_key is already taken into account by gr.spatial_neighbors (see also #917).

Am I missing something?