extend options to aggregate ranks across batches for HVG selection

I would like to get an opinion from @gokceneraslan, @adamgayoso (who implemented this originally), or someone else more familiar in hvg selection here.

I feel like we should probably have multiple methods for this, which we expose via a kwarg. I'm not confident median rank makes sense as a value here. I think there are assumptions about what kinds of values a rank can have that get broken by taking the median.

Is the method for aggregating hvgs from multiple batches here reasonable? For reference, a similar code block can be found in _highly_variable_genes_seurat_v3, which says this is the method used in SelectIntegrationFeatures from Seurat.

Originally posted by @ivirshup in https://github.com/theislab/scanpy/pull/1715#discussion_r736690390

@jlause commented:
I agree that exposing multiple options could make sense here. Any opinion @gokceneraslan and @adamgayoso which ones we should implement / which one should be the default?

I just copied this part from the seurat_v3 batch integration, so I don't have an expert opinion - but if I remember correctly, the other flavors use dispersions_norm to break ties instead of using the median rank. If I read the code correctly, this is the within-batch dispersion averaged over batches.

We could also do something similar here: Compute residual variance within each batch, take the average across batches, and use that to break ties instead of using median rank.

Start with the existing batch aggregation logic and compare it with the code block in _highly_variable_genes_seurat_v3, which follows Seurat's SelectIntegrationFeatures. Investigate the discussion around median rank, dispersions_norm, and averaged residual variance. Done means agreeing on supported aggregation methods, selecting a default, and exposing the choice through a keyword argument.