This is kind of a followup to astral-sh/ruff#21173 and pertains to behaviours described in imports/nonstandard_conventions.md
We do not currently consider from .a.b import c in an __init__.py(i) as ensuring thispackage.a.b or thispackage.a.b.c are submodule attributes in files that import thispackage, even though logically anything that acquires a reference to that module must have executed that import.
Ideally it should, but I think for safety it requires:
(This one may be a bad idea for performance/caching/isolation reasons, but I think it might be ok since we have to load and analyze that package's module anyway when you access an attribute on it?)