Load USEARCH matrix as scipy sparse matrix

[prihoda]

The sparse matrix is problematic since missing values have to be 0, which is not what we want from a distance matrix. A solution might be to use the exact UMAP metric.

[prihoda]

lmcinnes, is there more information about the exact metric? Is it already implemented? I see it in the docs, but not in the code. In this tool, I am actually loading the distances as a dense numpy matrix for the reason mentioned above.