fix #1822 regression from 1.3b3

[jschueller]

it reverts part of 21b8384 that was a backward-incompatible attempt to fix #1061.
Maybe this issue should be solved by using a backward-compatible :imported: flag as suggested by @shimizukawa, see #1822 for details

[embray]

👍 By the way there is already an option to ModuleDocumenter called "imported-members", which when set to True also works around this--it forces documenter.check_module() to return True. The problem is that there's no obvious place for autosummary to pass that option to autodoc.

I worked around the backward-incompatible change that this fixes in my subclass of Autosummary like this: https://github.com/astropy/astropy-helpers/pull/148/files#diff-c95430c65faa667beb1feea7abfb3fd6R188

A better longterm solution, as suggested, would be to add an equivalent flag directly to Autosummary as well.

[christianbrodbeck]

I would be happy with this fix, are there reasons it is held up?

[shimizukawa]

Sorry for long time holding (I have no time past 3 months..).
I'll take a look in a week.

[shoyer]

Not sure how much it helps to pile on, but IMO this is a major regression and it would be prudent to get a release with a fix out ASAP. This broke API docs for (at least) scikit-learn, pandas, xray and dask. The first two of those projects are foundational projects for the scientific Python stack.

[jschueller]

@shimizukawa, @Eric89GXL proposed a fix for this issue in #1892, maybe it is better, what do you think ?

[larsoner]

I tried to make the code accurately reflect what the documentation says it will do. I don't think this change does it as well as #1892 -- it does fix the problem, I remember having it as my first-order solution, but I can't remember the details of why #1892 seemed better. I think whatever rationale the person had for adding those two lines (that are deleted in this PR) made sense, but they needed the additional conditional to fit with the docs. And what the docs say make sense.

[jschueller]

agreed, refer to #1892