Add back getattr for ExtensionArrays

[dcherian]

cc @ilan-gold @martinfleis

[ilan-gold]

1. (fix): remove _getattr__ method for PandasExtensionArray #10250 (comment) should be considered here i.e., using __getattribute__ like I had had before the removal
2. @martinfleis PandasExtensionArray was also removed from the public API, but getattr was just a pass-through to the underlying array, which is now used in e.g., Variable.data instead of the PandasExtensionArray wrapper. So I would caution being 100% the two are linked.

[martinfleis]

I did not have time to pinpoint a specific commit but it has to be one of the commits merged in last 7 days, where this one felt like the only sensible option. But I might be wrong.

[ilan-gold]

@martinfleis I will have a look then! @dcherian What is xarray's breaking change policy? I was under the impression that there was none given the date-versioning, but should there be stronger guarantees?

[dcherian]

We try hard to not break things. Though in. this case, we're still figuring out how to handle extension arrays.

Seems like there are two options:

1. add back the passthrough
2. xvec extracts the underlying array.

I lean toward one, since that makes sense for a wrapper. I don't see any way other than __getattribute__ sadly, i thought __getattr__ would just work.

[ilan-gold]

@dcherian My caution/suspicion was warranted. I just checked out xvec and xarray before the removal of getattr. The test is still broken. What I did notice is that the dtype of df[c] on the breaking line is preserved now, which is more in line with the breaking change in: #9671. However, this is probably an upgrade for you @martinfleis - the xarray object now preserves the geometry data type because it was a coordinate and therefore under-the-hood, wrapped in an indexing adapter which #9671 addressed. So the conversion is likely no longer necessary.

[ilan-gold]

@dcherian As for bringing back this feature, I am neutral on it because it is not in the public API. I liked your suggestion of removing it because it means that things are explicit now. But of course, if you want to bring it back, __getattribute__ is a must.

@martinfleis I will open a PR

[ilan-gold]

@dcherian It would be great if we could merge this. We have lost the nbytes implementation so I think it might be only a matter of time until someone complains:

import geopandas, geodatasets

geopandas.read_file(geodatasets.get_path('geoda.malaria')).to_xarray()

[ilan-gold]

Suggested change

	return getattr(self.array, attr)
	return getattr(super().__getattribute__("array"), attr)