Add 40 complex PyMC model implementations

[twiecki]

Summary

Companion PR to #319. Adds PyMC translations for the remaining 40 Stan models that use advanced modeling features.

What makes these "complex"

• pm.Potential for Jacobian corrections and custom log-probability terms
• Helper functions for simplex transforms (ILR/Helmert), ODE solvers, GP spectral densities, HMM forward algorithms
• pytensor.scan for time series (GARCH, ARMA, state-space)

Models included

• IRT: 2pl_latent_reg_irt, grsm_latent_reg_irt, hier_2pl
• Capture-recapture: M0, Mb, Mh, Mt, Mtbh, Mth
• HMMs: hmm_drive_0, hmm_drive_1, hmm_example
• Topic models: ldaK2, ldaK5 (with ILR simplex transform)
• GPs: accel_gp, hierarchical_gp
• Time series: arma11, garch11, state_space_stochastic_level_stochastic_seasonal
• ODEs: lotka_volterra, one_comp_mm_elim_abs
• Forecasting: prophet
• Other: bones_model, bym2_offset_only, losscurve_sislob, multi_occupancy, nes, surgical_model, Survey_model, GLM_Binomial_model, and radon/logmesquite/Rate variants

Data wrangling

All numpy data wrangling occurs before with pm.Model(), per reviewer feedback on #319.

Test plan

• All 40 models pass gradient validation against BridgeStan (rtol=1e-5, atol=1e-6)
• All numpy operations moved before with pm.Model()
• Updated .info.json files to register PyMC implementations