Per categorical shuffling per dataset

[selmanozleyen]

Description of feature

• This can be either a new class or a new groupby parameter to the DatasetCollection API at this moment, I will leave this up to you @ilan-gold.
• Why would it be multiple classes if it's only one parameter? Idk, maybe to leave space to changes in case we want something more elaborate than one pass sorting.
• This basically just shuffles within categories per dataset inside.

[ilan-gold]

I think just adding a groupby parameter to the current DatasetCollection should suffice. As discussed, the Sampler can then do some sort of validation on min-chunk-size for categorical runs. The worst case is that it degenerates into perfect random sampling.

In the future, we can think about writing out metadata that indicates if a collection is group-by'ed or not. For now I don't see a need since we're not writing per-category datasets with this route.

[selmanozleyen]

So since the boundary data can be coupled with categorical data we might have encoding problems. So here are some options:

• Encode only the boundaries so it can be a simple json attr in zarr collection, the labels can be resolved from obs anyways
• Store it under uns so that we can store it as boundaries_table with proper categoricals
• Don't store under uns but use write_elem to store a table for it. Not on anndata level but collection level.

My gut says I should go for 1st option, wdyt?