Refactor shuffling functions to be methods on one class called `Shuffler`

[ilan-gold]

Description of feature

This pairs well with #93 so it will be clear we have two main imports, Batcher and Shuffler

Goals:

1. Class implementation that gives users an "object" to work with
2. Write out metadata that indicates a collection of anndatas was indeed preshuffled
3. A to_adata(lambda x: ad.AnnData(X=x.X, ...)) type function on the class so that you can quickly go from shuffled store (which has uniform keys, Xs etc.) to data loading fast (compare this to the current state where you have to manually open anndata files

[felix0097]

Does it make sense to use PreShuffler instead of Shuffler? @ilan-gold

[ilan-gold]

Things move fast in AI - my perspective has already shifted since writing this. I think a Collection object is actually the correct name/abstraction with the advantage being we could do Loader.add_collection(Collection(...)) (after shuffling bia Collection.shuffle(adatas) or something) instead of opening 50 anndata files manually. Or perhaps there is a Collection.to_adatas method. Let's talk tomorrow though.

[falexwolf]

I agree on Collection being a good abstraction ☺️

[ilan-gold]

After thinking things through, I think PreShuffler is best because of the anndata semantics - we need to use the anndata API in the loaders for handling obs and X etc. so we can't just blindly load a Collection