Zarr3 vs Zarr2 benchmarking

[anaprietonem]

Initial tests suggest that model performance is not impacted (in terms of loss function or skills) but we do see a slight drop in throughput and also increase in CPU memory usage. We want to run some bencharmking run to better understand that.

[frazane]

Just for background: https://github.com/zarrs/zarr_benchmarks

It might be interesting to consider zarrs-python.