Generalize preprocessing to handle custom chunking / `X` not as a dask Array

[ilan-gold]

Description of feature

See https://github.com/scverse/annbatch/pull/65/files#r2448590922

Our handling of individual anndata input objects relies on X being a dask array with a sensible chunking. But this is probably overly restricitive as

1. we could in theory handle in memory objects as well, in which case no chunking is needed (or maybe custom?).
2. a key in layers could have the chunking we're actually interested in (in fact, we should probably recommend people use layers for a variety of reason, not the least of which is explainability).