Opening large Zarr datasets should be lazy (and fast)

[alxmrs]

Found out when trying to open ARCO-ERA5. It looks like it eagerly converts data to a dask dataframe.

Instead, it should return immediately with deferred calls to convert the chunks to dfs.

Implementation Notes

Change from_map to from_delayed.

[alxmrs]

https://docs.dask.org/en/latest/delayed-collections.html