Add a type guard for intX

[ferrine]

• what are the (breaking) changes that this PR makes?
  Integer dtypes are now not downcasted in intX
• important background, or details about the implementation
  In discrete distributions there is a usage of intX that enforces unreasonable limitations in UX (see Do not force lower precision in intX #4553)
• are the changes—especially new features—covered by tests and docstrings?
  I did not yet create a unit test covering this specific usecase
• linting/style checks have been run
  Ran black on it
• consider adding/updating relevant example notebooks
  Not required
• right before it's ready to merge, mention the PR in the RELEASE-NOTES.md

Note: Questions before we go ahead, the alternative would be leave the function as is and covert parameters in Discrete distributions

[ferrine]

@lucianopaz pymc3/tests/test_model_helpers.py::TestHelperFunc::test_pandas_to_array this test does not pass. I've changed intX to be tolerant to int dtypes but it complains about new behavior: not downcasting int32 to int16. What is treated as the expected behavior here?

[ferrine]

also, @brandonwillard, do we really need downcasting there?

[lucianopaz]

Sorry for taking so long @ferrine. I honestly don’t know what’s best here. Given that you’ve changed the intX behaviour it would make sense to also change the test function to adapt to the new behaviour. I imagine that you would have to change the if statement to make it more lenient with ints.

[ferrine]

I'll fix the test soon as it seems we all agree on the change

[ferrine]

I've fixed the failings tests, feel free to review and merge

[ferrine]

@lucianopaz or @brandonwillard can you please have a look at this?

[michaelosthege]

@ferrine not converting already int data sounds like a good idea. And your solution is simple yet elegant.
Go ahead and merge unless you changed your mind

[ricardoV94]

What about pushing this to the V4?

[twiecki]

Yeah, we should.