Dataset cannot be opened when underlying `zarr` arrays have `fill_value` set.

[benjeffery]

Due to pydata/xarray#7292 zarr arrays with fill_value set get their dtype changed to float32. This causes the dataset to fail when opened with sgkit.

fill_value is used in zarr to set a default value for arrays, this can be useful as it allows chunks that only contain the default value to not be written to disk at all. I was investigating using this for tskit-dev/tsinfer#729 where we are storing many short haplotypes.