`hdf5` support via `virtualizarr`

[ilan-gold]

Description of feature

Steps:

1. Custom HDFParser from virtualizarr for full anndata objects (encoding object semantics like string arrays etc.). Also ensure we have a way to create icechunk stores from this parser (maybe doesn't exist in virtualizarr given how xarray focus the packages is)
2. Integrate icechunk store support into zarrs-python so we can read the virtualized output of step 1
3. Add API here for handling these stores, both in preshuffling and data loading

[ilan-gold]

We just had a user try shuffling the entire CellXGene corpus and doing this, even without zarrs, was a massive speedup for reshuffling. In fact, it's basically necessary at least on Helmholtz's filesystem to not use HDF5 reader software.

I think we can get away with custom parsers for just array data and worry about object types at a later date.