Finding a better home for cluster centroids (i.e. 🚫_tmp_cluster_pos)

We have a weird temporary global variable called sc.pl._utils._tmp_cluster_pos. We use it for storing the positions of cluster centroids (actually the centroids of any categorical variable for any sort of embedding). The weird part is that it's set in scatterplot functions (see https://github.com/theislab/scanpy/blob/master/scanpy/plotting/_anndata.py#L468 and https://github.com/theislab/scanpy/blob/master/scanpy/plotting/_tools/scatterplots.py#L809) and used only by sc.pl.paga_compare (https://github.com/theislab/scanpy/blob/master/scanpy/plotting/_tools/paga.py#L119).

First, it's not obvious where paga_compare finds centroids (it was a mystery to me until recently). Second, the current design is error-prone (see a corner case https://github.com/theislab/scanpy/issues/686). Therefore, there should be a better place to store cluster centroids :)

I'm not following the discussion about the future of AnnData, but maybe having something like adata.uns['obs_category_leiden'] and storing colors and centroids in it e.g. adata.uns['obs_category_leiden']['colors'] and adata.uns['obs_category_leiden']['centroids']['X_umap'] would be more structured.

Read the scatterplot code in scanpy/plotting/_anndata.py and scanpy/plotting/_tools/scatterplots.py where _tmp_cluster_pos is set, then inspect scanpy/plotting/_tools/paga.py where it is consumed. Review issue #686 for the error-prone corner case. Done means paga_compare obtains centroids from a structured, reliable location rather than the temporary global.