Fixed bug in delta_logp for MLDA that broke AEM and VR

[mikkelbue]

This PR fixes #5099

The bug was caused by the now deprecated function delta_logp_inverse not evaluating the proposal last, which was the sole intention with that function. The functionality has now been replaced by

• using a lambda-function for the base sampler when using VR, where additional information is collected from the base model.
• swapping q and q0 and the sign of delta_logp for the finer sampling levels.

[codecov]

Codecov Report

Merging #5104 (fbcdecc) into main (38295b7) will increase coverage by 9.90%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main    #5104      +/-   ##
==========================================
+ Coverage   67.91%   77.81%   +9.90%     
==========================================
  Files          87       87              
  Lines       14024    14006      -18     
==========================================
+ Hits         9524    10899    +1375     
+ Misses       4500     3107    -1393     

Impacted Files	Coverage Δ
pymc/step_methods/mlda.py	96.40% <100.00%> (+45.61%)	⬆️
pymc/backends/base.py	86.25% <0.00%> (+0.76%)	⬆️
pymc/step_methods/hmc/base_hmc.py	91.12% <0.00%> (+0.80%)	⬆️
pymc/step_methods/hmc/hmc.py	92.30% <0.00%> (+1.92%)	⬆️
pymc/backends/report.py	89.51% <0.00%> (+2.09%)	⬆️
pymc/distributions/discrete.py	98.02% <0.00%> (+2.30%)	⬆️
pymc/util.py	74.35% <0.00%> (+2.56%)	⬆️
pymc/distributions/distribution.py	94.56% <0.00%> (+2.71%)	⬆️
pymc/step_methods/slicer.py	95.71% <0.00%> (+2.85%)	⬆️
pymc/distributions/continuous.py	95.59% <0.00%> (+3.14%)	⬆️
... and 18 more

[twiecki]

Thanks!