Systematized Distribution testing overlooks Theano arguments

[brandonwillard]

I noticed that our testing framework doesn't perform tests using Theano arguments (not within our systematized tests—like TestMatchesScipy—at least).

This seems like a pretty big oversight, and appears to have let #3999 go unchecked.

In general, I would say that Theano arguments make a much better default than NumPy, so perhaps we should consider changing our test framework to use them.

Originally posted by @brandonwillard in #4000 (comment)

[ricardoV94]

@brandonwillard I think the main check_logp and check_logcdf (the later after #4736) in TestMatchesScipy are now testing with aesara arguments via shared variables. Should we close this one?

[brandonwillard]

If we're talking about v4, then the concerns expressed in this issue aren't really the same/present. This issue was prompted by errors in our our mixed use of symbolic/non-symbolic values (i.e. special handling for values that could be either type).