distances are buggy, should be re-exports from pynndescent

[ilia-kats]

The umap package seems to reimplement the same distance metrices as the pynndescent package. However, the umap implementation is buggy in at least one case:

• sparse.arr_union can return one of its input arrays
• the return value of sparse.arr_union is used as a writable buffer by sparse.sparse_sum
• sparse.sparse_diff is a thin wrapper around sparse.sparse_sum
• sparse.sparse_diff is used in sparse_euclidean

This leads to two issues:

1. The input array may be modified, which is not expected by the user
2. If sparse_euclidean, sparse_diff, or sparse_sum is called by a custom distance metric, which gets its sparse array via keyword arguments, the keword arguments will be stored in a closure, which leads numba to treat them as readonly, leading to type inference failure (see Cannot call neighbors: Failed in nopython mode pipeline (step: nopython frontend) scverse/muon#173).

The pynndescent implementation of sparse_euclidean works without any issues. Given that both packages are maintained by you, umap depends on and uses pynndescent, and the supported distance functions appear to be identical, I think it would make sense if umap would simply re-export the distance functions from pynndescent.