Remove all uses of *.tag.test_value and replace with a new initial value framework

[brandonwillard]

We need to set up a framework for setting and getting initial values for a model; one that doesn't rely on Aesara's test values.

In many cases, this may be as simple as renaming parts of the existing API (e.g. Model.test_point to Model.initial_point, Model.check_test_point to Model.check_initial_point, change the testval Distribution argument to initial_value, etc.), but, in some cases, we'll need to add new variables and logic.

Specifically, it seems like we'll need to store the initial values in the Model object and, in multiple places, obtain those values from a Model object.

[farhanreynaldo]

I would like to take on this issue. Should I do the pull request on v4 branch? Would also really appreciate feedback and your help on this issue

[brandonwillard]

You might need to wait until #4577 goes through, because it looks like I'll need to implement some of this for other reasons.

[farhanreynaldo]

Sure thing, I will start working when it's merged.

[ricardoV94]

For whomever takes this one, the Flat and HalfFlat are examples of pymc3 "distributions" that require special care because they cannot be sampled from directly, and as such an initial value has to be chosen some other way.

[ricardoV94]

@brandonwillard can we close this one?

[twiecki]

Sounds like this is implemented now.