Fix AttributeError in underline length check

[rossbar]

Fixes gh-362.

The NumpyDocString class raises a UserWarning during docstring parsing when the underline length doesn't match the header length. In an attempt to make this UserWarning more informative, some introspection is performed to try to get the name and location of the object. This introspection relied on the assumption that the object would have a __name__ attribute, which is not always the case, e.g. when _obj is an instance of a user-defined Callable.

This PR fixes the issue by only reporting the improved UserWarning message when the introspection is successful.

[larsoner]

Thanks @rossbar !

[ajschmidt8]

@larsoner, will this be included in a release soon? This fixes an error that we're seeing with our doc builds.

[rossbar]

will this be included in a release soon? This fixes an error that we're seeing with our doc builds.

I'd be +1 for a patch release but we might want to wait a bit to see if any more bugs come up.

Fortunately, this particular exception is triggered by a Warning due to bad underline length for section headings, so fixing the docstring heading issue will also fix the exception. See e.g. statsmodels/statsmodels#8044, which was all that was required to fix the error that was originally reported in #362.

Your case should be similar; the trick will be finding which docstring hit the exception. If you look at the sphinx-build log, you should see something like:

Handler <function mangle_signature at 0x7f0833af8f70> for event 'autodoc-process-signature' threw an exception (exception: '<object_name>' object has no attribute '__name__')

or similar. If you look at the docstring for <object_name> you should be able to find the docstring section heading(s) with mismatched underline length. Fixing the heading underlining should be sufficient to avoid the exception.