batch integration baselines are suboptimal

[scottgigante-immunai]

isolated label silhouette in batch_integration_embed and ARI in batch_integration_graph have scores much greater than 1. We need to design better baselines that are optimised to perform on these metrics. cc @LuckyMD

[LuckyMD]

may be solved via #686 . To be closed once results come out

[LuckyMD]

The iso label ASW baseline improvement was not solved via #686 for the pancreas dataset. It seems that the issue may be related to the isolated label being t cells in that dataset, which only has 7 cells. We should put a filter on n_cells per label in the dataset function to avoid this @danielStrobl. Maybe that could help this situation?

[LuckyMD]

See also #757

[LuckyMD]

@scottgigante-immunai... something strange is going on, as the baseline is perfect in scib. See theislab/scib#362

[scottgigante-immunai]

silhouette_batch still not resolved here.

[LuckyMD]

Maybe this should go to a different issue, as this issue was for iso label ASW and ARI, which are both addressed.