Check that sampler vars correspond to value variables in the model

[ricardoV94]

Closes #6223

I decided to not allow the old V3 behavior of passing intermediate variables/ Deterministic because I think that does something strange. Those variables are not the ones being sampled/maximized but instead the underlying input value variables are.

If a user wants to fit the variables that underlie a given graph, but he doesn't know what those are, they should do something like pm.find_MAP(vars=pm.inputvars(pm.aesaraf.rvs_to_value_vars(graph)))

Update: It is now allowed but a warning is emitted

Major / Breaking Changes

• ...

Bugfixes / New features

• ...

Docs / Maintenance

• Check that only random variables are assigned to samplers and find_MAP

[ricardoV94]

Why did the github bot remove the maintenance label?

[codecov]

Codecov Report

Merging #6235 (f66f320) into main (b5db350) will increase coverage by 2.32%.
The diff coverage is 94.59%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6235      +/-   ##
==========================================
+ Coverage   88.15%   90.47%   +2.32%     
==========================================
  Files         111      111              
  Lines       23820    23809      -11     
==========================================
+ Hits        20998    21542     +544     
+ Misses       2822     2267     -555     

Impacted Files	Coverage Δ
pymc/aesaraf.py	94.04% <ø> (ø)
pymc/step_methods/metropolis.py	83.36% <85.71%> (-0.11%)	⬇️
pymc/tuning/starting.py	91.73% <91.66%> (-0.77%)	⬇️
pymc/model.py	89.53% <100.00%> (+0.29%)	⬆️
pymc/smc/kernels.py	97.37% <100.00%> (ø)
pymc/step_methods/hmc/base_hmc.py	89.90% <100.00%> (+0.09%)	⬆️
pymc/step_methods/slicer.py	96.15% <100.00%> (-0.05%)	⬇️
pymc/tests/tuning/test_starting.py	100.00% <100.00%> (ø)
pymc/backends/arviz.py	90.08% <0.00%> (+2.89%)	⬆️
... and 3 more

[ricardoV94]

This recent Discourse topic touches on the same issue of trying to pass something that's not an RV to a function that expects RVs: https://discourse.pymc.io/t/trying-to-use-pm-binarymetropolis/10645

With the changes in the PR the new error message would be (condensed example):

import pymc as pm
with pm.Model() as logistic_model_pred:
    p = pm.Beta("p", 1, 1)
    label = pm.Data("label", value=[0, 1, 2], mutable=True)
    observed = pm.Bernoulli("occupancy", p, observed=label)
    step = pm.BinaryMetropolis(vars=label)    

Traceback (most recent call last):
  File "/home/ricardo/miniconda3/envs/pymc/lib/python3.10/site-packages/IPython/core/interactiveshell.py", line 3397, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-2-f13ba6f7e166>", line 6, in <cell line: 2>
    step = pm.BinaryMetropolis(vars=label)
  File "/home/ricardo/Documents/Projects/pymc/pymc/step_methods/metropolis.py", line 395, in __init__
    vars = get_value_vars_from_user_vars(vars, model)
  File "/home/ricardo/Documents/Projects/pymc/pymc/util.py", line 403, in get_value_vars_from_user_vars
    raise ValueError(
ValueError: The following variables are not random variables in the model: ['label']

[ricardoV94]

Ready for review

[junpenglao]

LGTM. Minor nit but feel free to ignore.

[ricardoV94]

The MacOS job got stuck in conda-cache. Gonna take a risk and merge regardless