This addresses several regressions identified in #16486

The primary regression from #15770 is that pybind11 properties with docstrings were erroneously assigned typeshed. Incomplete.

The reason for the regression is that as of the introduction of the --include-docstring feature (#13284, not my PR, ftr), ./misc/test-stubgenc.sh began always reporting success. That has been fixed.

It was also pointed out that --include-docstring does not work for C-extensions. This was not actually a regression as it turns out this feature was never implemented for C-extensions (though the tests suggested it had been), but luckily my efforts to unify the pure-python and C-extension code-paths made fixing this super easy (barely an inconvenience)! So that is working now.

I added back the extended list of typing objects that generate implicit imports for the inspection-based stub generator. I originally removed these because I encountered an issue generating stubs for PySide2 (and another internal library) where there was an object with the same name as one of the typing objects and the auto-import created broken stubs. I felt somewhat justified in this decision as there was a straightforward solution -- e.g. use list or typing.List instead of List. That said, I recognize that the problem that I encountered is more niche than the general desire to add import statements for typing objects, so I've changed the behavior back for now, with the intention to eventually add a flag to control this behavior.