Fix exponential and gamma logp / random link

[ricardoV94]

After merging #4548, the Exponential and Gamma distributions still have a mismatch between the logp and random parametrizations.

I still think it is more straightforward to keep relying on the tests in test_distributions_random.py while we refactor the distributions into V4 as it easily identifies this type of issues. There is already an open issue to refactor these tests #4554, so it will not be forgotten.

For the discussion into the Exponential and Gamma changes see my previous comments:

1. Exponential: Fix different / alternative parametrizations between RandomOps and Logp methods #4548 (comment)
2. Gamma: Fix different / alternative parametrizations between RandomOps and Logp methods #4548 (comment)

There seems to be a problem with the Categorical (test_distributions_random::TestScalarParameterSamples::test_categorical_random), but perhaps it was addressed by pymc-devs/pytensor#351?

[brandonwillard]

We do not need to create that new RandomVariable or add a new optional parameterization at this time.

[ricardoV94]

What do you suggest instead? Shall we leave the distributions not matching parameters on the logp and random sides?

For the exponential even if we don't provide it as an option, we still need to use mu behind the scenes since that's how the aesara op is parametrized.

The gamma right now requires changing the aesara op one way or another as explained in the comment above.

[brandonwillard]

What do you suggest instead? Shall we leave the distributions not matching parameters on the logp and random sides?

We only need to keep the Distribution interface, so, inside Exponential.dist, we can at.inv the parameter and update the corresponding logp and logcdf functions. The latter two functions are not called directly, so they only need to correspond to the RandomVariable parameterizations.

[ricardoV94]

I removed the explicit new parametrization of the exponential and forced the gamma to work with the parameter inversion on the aesara side. The last one still feels weird, but if that's how we want to keep it, so be it.

[brandonwillard]

I rebased this branch onto the update v4.

[ricardoV94]

Thanks!

[ricardoV94]

I changed the exponential equal assertion to almost_equal as it was failing on float32. The tests should pass now.

[ricardoV94]

Unrelated to this PR this test failed:
test_sampling.py::test_exec_nuts_init