Possibility of using chem_wasserstein as a proper UMAP metric

[ml-evs]

It is possible to use custom metrics in UMAP, provided they are numba njit'd all the way down. This enables the metric to be used with umap-learn's built-in nearest neighbor implementation, which allows for new data to be embedded without requiring the full distance matrix computation.

I've tried several modifications of EMD etc. to attempt to enable this (requires pre-computing the ratio_vectors, but then I end up in a seemingly infinite stack of numba errors inside the network simplex algorithm). Is there any interest in supporting this within this package directly? It's quite hard to follow the levels of indirection between ElMD, chem_wasserstein, dist_matrix etc. here so would appreciate it if you already know if this would be hard/easy to extract out. Cheers!