TST: DocTestFailure 'MaskedArray' object has no attribute 'unit'

Hold your horses, @astrofrog and @WilliamJamieson ! We need to fix this before merging more black stuff. I think one of them broke a doctest. The PRs might be fine individually but maybe didn’t play well together. cc @mhvk

Example log: https://github.com/astropy/astropy/actions/runs/3465731975/jobs/5788828409

_____________________________ [doctest] index.rst ______________________________
119 |Quantity|, but does not share any of its methods.  Hence, even though the
120 resulting class looks reasonable at first glance, it does not work as expected::
121 
122   >>> q = [1., 2.] * u.m
123   >>> np_mq = np.ma.MaskedArray(q, mask=[False, True])
124   >>> np_mq
125   masked_Quantity(data=[1.0, --],
126                   mask=[False,  True],
127             fill_value=1e+20)
128   >>> np_mq.unit
Differences (unified diff with -expected +actual):
    @@ -1,3 +1,6 @@
     Traceback (most recent call last):
    -...
    -AttributeError: 'MaskedArray' object has no attribute 'unit'
    +  File "/opt/hostedtoolcache/Python/3.10.8/x64/lib/python3.10/doctest.py", line 1350, in __run
    +    exec(compile(example.source, filename, "single",
    +  File "<doctest index.rst[21]>", line 1, in <module>
    +    np_mq.unit
    +AttributeError: 'MaskedArray' object has no attribute 'unit'. Did you mean: '_unit'?

/home/runner/work/astropy/astropy/docs/utils/masked/index.rst:128: DocTestFailure

[cfbolz]

or you could do this kind of thing:

    >>> foo.bar # doctest: +ELLIPSIS
    Traceback (most recent call last):
    ...
    AttributeError: 'Foo' object has no attribute 'bar'...

note the ... at the end. this still checks the important part of the message but ignores the suggestions.

[cfbolz]

I don’t know yet why stuff breaks with pytest, if exceptiongroup is to blame I’m happy to look into that more. Just chiming in to say that it’s very much intentional on the CPython side that the suggestions are only computed when printing the traceback. computing the suggestion is relatively costly (it does dir(obj), then computes an edit distance between the wrong attribute and all other attributes) so it would be way too expensive to do that computation on every instantiation of AttributeError, some of which might be caught and never even reach the user.