Unseed (and slightly refactor) TestMatchesScipy

[ricardoV94]

This PR unseeds TestMatchesScipy, which contradicted the effect of limiting the number of samples in product(domains, n_samples) which is supposed to return a random subset of domain x paramdomain points that are used to test the logp and logcdf methods.

This, for instance, made this last PR #4459 somewhat shortsighted in that it assumed we would keep covering the entire domain grid (over multiple CI runs)

I also refactored the old pymc3_matches_scipy to check_logp, in line with the check_logcdf that was added at a later point. I added some documentation, mostly borrowing from the check_locgdf function.

---

Depending on what your PR does, here are a few things you might want to address in the description:

• what are the (breaking) changes that this PR makes? None
• important background, or details about the implementation. None
• are the changes—especially new features—covered by tests and docstrings? New docstrings
• linting/style checks have been run. Yes
• consider adding/updating relevant example notebooks. NA
• right before it's ready to merge, mention the PR in the RELEASE-NOTES.md. Internal issue, NA

[ricardoV94]

Seems like we have a numerical precision issue in the skewnormal: https://github.com/pymc-devs/pymc3/pull/4461/checks?check_run_id=1837831640

Reducing the 64bit decimal precision from 6 to 5, seems to work (I tested on all combinations locally)

[ricardoV94]

Precision issues in gamma logcdf (32-bit): {'alpha': array(20., dtype=float32), 'beta': array(0.9, dtype=float32), 'value': array(0.01, dtype=float32)} -> -inf in the PyMC3 implementation

[ricardoV94]

InverseGamma logp also failed with 6 decimals, but passes with 5 (all combinations tested locally).

I am now testing all methods locally with all the samples (64bit for now). I figured out that it's probably better to do a complete run once rather than keep chasing the failures...

[ricardoV94]

My only lingering concern is whether check_int_to_1 has any randomness to it, which could benefit from seeding. Does anyone know? If yes, I would move the 7 or so tests that use it a separate Seeded class.

https://github.com/pymc-devs/pymc3/blob/60e0daea62f1d3a024bc62aca6703ce0eafbdea5/pymc3/tests/test_distributions.py#L735

[michaelosthege]

My only lingering concern is whether check_int_to_1 has any randomness to it, which could benefit from seeding. Does anyone know? If yes, I would move the 7 or so tests that use it a separate Seeded class.

https://github.com/pymc-devs/pymc3/blob/60e0daea62f1d3a024bc62aca6703ce0eafbdea5/pymc3/tests/test_distributions.py#L735

integrate_nd uses scipy.integrate.quad which should be deterministic as far as I know.

[michaelosthege]

LGTM