Presence of empty `__init__.py` breaks test discovery in hard-to-diagnose ways

Hey all

I'm trying to add another package to conda-forge, and ran into collection problems when attempting to run the upstream test suite:

==================================== ERRORS ====================================
______ ERROR collecting tensorflow_addons/callbacks/tests/run_all_test.py ______
import file mismatch:
imported module 'run_all_test' has this __file__ attribute:
  $SRC_DIR/tensorflow_addons/activations/tests/run_all_test.py
which is not the same as the test file we want to collect:
  $SRC_DIR/tensorflow_addons/callbacks/tests/run_all_test.py
HINT: remove __pycache__ / .pyc files and/or use a unique basename for your test file modules

The first thing that caught my eye were the weird kind of test wrappers that tf-addons is using (my guess from searching their repo is that this is for bazel compatibility, and hence not so relevant due to the relative rarity of that in python; edit: beyond the convoluted definition, the underlying issue for these is #9301), but removing those still left me with similarly cryptic errors:

==================================== ERRORS ====================================
_ ERROR collecting tensorflow_addons/callbacks/tests/avg_model_checkpoint_test.py _
ImportError while importing test module '$SRC_DIR/tensorflow_addons/callbacks/tests/avg_model_checkpoint_test.py'.
Hint: make sure your test modules/packages have valid Python names.
Traceback:
../[...]/lib/python3.9/importlib/__init__.py:127: in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
E   ModuleNotFoundError: No module named 'tests.avg_model_checkpoint_test'

I then got curious about No module named 'tests.avg_model_checkpoint_test', especially because the upstream test modules (an example) looked alright, and that module that was supposedly not being found didn't appeared anywhere.

By chance, I noticed that the test folders contain empty __init__.py files, and due to some dim awareness of the issues around imports, I decided on a whim to see what happens if I remove them - turns out collection succeeds then! 🥳

From looking around the issue tracker (and said awareness of import troubles in python), I believe this issue is probably quite thorny, so the underlying problem is probably captured in other issues already: #3396 #5147 #8332

However, I'm hoping that a possible fix for empty __init__.py might be easier than that, and at the very least, I think the warning should be improved, because it's really not obvious what's causing the problem.

Reproduce collection failures with empty init.py files in the test folders, using the tensorflow-addons test paths and errors described here. Compare collection after removing those files, then inspect pytest's test discovery and import handling; done means the failure is fixed or the warning clearly identifies the cause.