I feel like it would be good to have an option to switch between zarr/zarrita/tensorstore for the zarr i/o, as they all have pros and cons and this would enable benchmarking each of them.

The most useful think would probably to abstract away the few features that we actually use (opening/writing a group, opening/writing/slicing an array), and wrap each of the libraries in classes that inherit from this AbstractZarrIO class.