Add 68 idiomatic PyMC model implementations

[twiecki]

Summary

• Add PyMC translations for 68 Stan models that follow idiomatic PyMC patterns
• No pm.Potential, no custom helper functions
• All numpy data wrangling occurs before the pm.Model context manager
• Updated .info.json files to register PyMC implementations

Model categories included

• GLMs: wells (8 variants), radon (10 variants), kidscore (7 variants), logearn (4 variants), logmesquite (2 variants)
• Gaussian Processes: gp_regr, gp_pois_regr, accel_splines
• Mixture models: normal_mixture, normal_mixture_k, low_dim_gauss_mix (2 variants)
• IRT: irt_2pl
• Hierarchical: election88_full, GLMM1_model, GLMM_Poisson_model, GLM_Poisson_model, pilots
• Educational: eight_schools_centered, rats_model, diamonds, blr
• Ecological: seeds (3 variants), dogs (2 variants), dugongs_model
• Other: arK, logistic_regression_rhs, Rate models, mesquite, sesame_one_pred_a, etc.

What is NOT included (for separate PR)

• Models using pm.Potential (40 models) - need review of whether Potentials can be eliminated
• Models with custom helper functions (prophet, ODE models, simplex models, etc.)

Test plan

• All models verified gradient-equivalent to Stan via BridgeStan (rtol=1e-5, atol=1e-6)
• Zero pm.Potential usage across all 68 models
• Zero helper function definitions (only make_model)
• All numpy operations occur before with pm.Model()

[ricardoV94]

Looking pretty nice, left some comments that apply more in general than the places I noted.

Big issues:

• Lot's of dangling commets/logic for "normalization correction/matching Stan" that is not being done anymore
• Way too many dumb comments in general, let the code speak unless it really seems pertinent
• Some models have the prior_only but most don't? Is this a thing in the original posteriordb or the LLM just forgot about it?
• Common to see the same expression defined once in the likelihood and then repeated inside a Deterministic. Not idiomatic at all

If we fix those, this is pretty much ready. Can you push a commit on top of this, instead of regenerating everything and force-pushing? Or even if you regenerate everything, don't force push

[twiecki]

@ricardoV94 should be addressed.

[ricardoV94]

LGTM

[twiecki]

@MansMeg this should be good to go.

[JTorgander]

Big thanks for the contribution!

I’ll run the models through our internal test suite, which appears very similar to what you’ve been using. One initial observation:

Importing the Python modules (corresponding to the PyMC models) raises a NameError on the return type hint:
def make_model(...) -> pm.Model:

This happens because pm is not defined at function definition time. I suggest moving the module imports (e.g., import pymc as pm) to the module level rather than inside the make_model functions.

[JTorgander]

Please revert this change here and in subsequent files.