Do not rely on tag information for rv and logp conversions

[ricardoV94]

This PR will finally clean the interface for accessing meta-information from RVs needed when transforming back and forth between the RV and logp representations of a PyMC model

It does not yet remove that information to not break anything. Some methods relied heavily on this local information and were simply marked for deprecation, with alternatives being provided from scratch.

Accessing these variables (as long as they were introduced by PyMC) via tag will issue a FutureWarning:

with pm.Model() as m:
  x = pm.Normal("x")
x.tag.value_var  # Warning: The tag attribute value_var is deprecated. Use model.rvs_to_values[rv] instead

Closes #5033
Closes #6236

Major / Breaking Changes

• Deprecated accessing [value_variable|observations|transform|total_size] via var.tag in favor of model.rvs_to_[values|transforms|total_sizes]
• Deprecated joint_logp in favor of model.logp
• Deprecated aesaraf.rvs_to_value_vars in favor of model.replace_rvs_by_values

Bugfixes / New features

• Fix bug when replacing random variables with nested value transforms

Docs / Maintenance

• ...

[codecov]

Codecov Report

Merging #6281 (2b266dc) into main (0be5295) will decrease coverage by 0.18%.
The diff coverage is 97.56%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6281      +/-   ##
==========================================
- Coverage   94.30%   94.12%   -0.19%     
==========================================
  Files         111      111              
  Lines       23809    23861      +52     
==========================================
+ Hits        22452    22458       +6     
- Misses       1357     1403      +46     

Impacted Files	Coverage Δ
pymc/backends/arviz.py	90.08% <50.00%> (ø)
pymc/tests/backends/fixtures.py	78.30% <50.00%> (ø)
pymc/distributions/logprob.py	67.28% <92.59%> (-29.98%)	⬇️
pymc/model.py	89.68% <96.29%> (+0.15%)	⬆️
pymc/aesaraf.py	94.86% <100.00%> (+0.82%)	⬆️
pymc/distributions/distribution.py	95.02% <100.00%> (+0.02%)	⬆️
pymc/distributions/shape_utils.py	97.87% <100.00%> (+0.01%)	⬆️
pymc/initial_point.py	100.00% <100.00%> (ø)
pymc/model_graph.py	78.82% <100.00%> (ø)
pymc/sampling/forward.py	95.55% <100.00%> (ø)
... and 15 more

[michaelosthege]

@ricardoV94 the type hint in aeppl.transforms.TransformValuesRewrite is wrong: https://github.com/aesara-devs/aeppl/blob/961496a17e4291b76a5affc140ce85b42a6527cb/aeppl/transforms.py#L193

It should say Dict[TensorVariable, Optional[Union[RVTransform, DefaultTransformSentinel]]]

Best place a # type: ignore in the line here for now

[michaelosthege]

Overall looking good; great work, @ricardoV94 👍

I found some minor things in docstrings and I'm slightly worried that the big multiline expressions in some tests might result in hard to read tracebacks if they fail but this is also just a minor concern.

[ricardoV94]

I'm slightly worried that the big multiline expressions in some tests might result in hard to read tracebacks if they fail but this is also just a minor concern.

Then we will make sure they never fail :D

Yeah those tests a bit messy, but on the bright sight they are due to be removed when we deprecate the warnings.