A recent unit test was affected by a bug in TestModelsLoaderMixin

causing a models module to remain in Django's cache after the test suite

concluded. This only happened if a subclass defined a tests app that

already contained a models.py file. If it didn't define a specific

tests app or didn't include a models.py, then the cleanup would work

correctly, but if both were present then it would fail.
The reason for this inconsistency is that if a valid models.py didn't

exist, we'd create one and store a reference to it, which allowed us to

later remove it from the module cache. If one already existed, then

there wouldn't be a reference, so it never got removed.
The logic for this is now a bit more careful to explicitly check for a

models module and to store it. If we can't import, we then proceed to

create a module and store that. This ensures we always have a module to

work with and to clean up at the end, fixing all tests.

All unit tests pass. Before this change, some failed with unknown app

errors involving models from prior test suites.