Recommendations for Zarr readers

[langmore]

I'd recommend a couple changes

• Change xr.open_zarr(path) to xr.open_zarr(path, chunks=None). Since you're not using Dask, this will speed things up. Without it, datasets with millions of chunks will be super slow to slice...as Dask will try to rebuild a graph (that you'll never use). It also blows up memory. FWIW, xbeam.open_zarr already passes chunks=None.
• The ARCO-ERA5 dataset is awesome, but they named it ....zarr-v3 (even though it's Zarr2). Your Xarray readers don't open this correctly because they try to auto-detect the filename ending. There are multiple ways around this.

BTW, I would just send a PR, but my previous PRs often sit around unsubmitted...so I figure it's easier for you to do it internally.