Cell2location model that considers cell type hierarchy when mapping

[vitkl]

This version of the model partitions cell abundance of broad cell types into sub-clusters - thus encoding the dependency that gene markers of a parent class (broad cell type) should be expressed in order for the model to detect child class in this location. E.i. if a location has 0 abundance of T-cells, it will have 0 abundance of T-cell subtypes.

[vitkl]

@la-sekretar this brach will eventually have the cell2location version that we discussed (that uses Beta/Dirichlet distributions to partition probabilities) and that needs signatures decomposed by https://cell2location.readthedocs.io/en/latest/cell2location.cluster_averages.html#module-cell2location.cluster_averages.markers_by_hierarhy