Label projection != label transfer

[rcannood]

This benchmark does not benchmark a method's performance on labelling a totally new dataset, given an existing reference dataset. Instead, it takes a large dataset and then uses one batch as test holdout.

• We should create a new task task_label_transfer which maps cell type labels from a reference dataset to a query dataset. This benchmark will then also have to deal with having to match different matrix shapes.
• To avoid confusion, we should update the description of this task to reflect the distinction between label projection and label transfer

[mxposed]

Hi Robrecht! Your comment brings up two questions:

1. This task was set up like this, because the correct labels on the test part were easily available. How would we get ground truth (or just correct labels) for label_transfer task between two datasets?
2. When transferring labels from one dataset to another, there can be a situation where some (or all) query cell types are not represented in the reference. I was thinking of emulating this by withholding some cell types in label_projection task. How do you think this should be handled in the label_transfer?