Add type annotations to `convert_observed_data`

[michaelosthege]

We use this function to typecast a zoo of different objects. When using it, it is completely unclear what you'll get back..

[ricardoV94]

The name should also be changed as it is no longer pandas specific (eg it has special functionality for masked arrays)

[MarcoGorelli]

https://github.com/pymc-devs/pymc3/blob/7e35cddd7e54e6e6f51fc1aa3ed3df76016aa00a/pymc3/aesaraf.py#L89-L90

What inputs could end up there, other than pandas objects?

[michaelosthege]

Someone could implement their own type with a to_numpy function I guess? At least there are no checks that prevent that from happening.

Probably more of a theoretical possibility though, but hasattr leaves that door open.

[MarcoGorelli]

True, but then you're assuming that their implementation of isnull returns either a numpy array or that it returns an object with to_numpy

FWIW I think just checking

if isinstance(data, (pd.Series, pd.DataFrame)):
    ...
elif isinstance(data, pd.Index):
    ...

could lead to fewer surprises

[michaelosthege]

If you promise that pandas won't remove the to_numpy and isnull methods in the future that's fine ;)

And it could help mypy so there's that.

[michaelosthege]

The function is being renamed by #5743.

[fonnesbeck]

It appears that convert_data (the new name) now has a return type.

[michaelosthege]

closed by #7299 and #7334