Bring back BART to V4 and make it more general

[aloctavodia]

This PR brings BART to V4, including some of the features in V3 (unreleased) plus a couple of changes not in V3 in order to make BART more general. For example It should now be possible to create semi-parametric models like:

with pm.Model() as model:
    μ = pm.BART("μ", X, Y)
    α = pm.Normal("α", 0, 1)
    β = pm.Normal("β", 0, 1)
    σ = pm.HalfNormal('σ', 1)
    y = pm.Normal("y", μ+α+β*X, σ, observed=Y)

There was a broadcasting problem when mixing NUTS and PGBART, that was actually a bug in base_hmc. It was finally fixed in #4952

[codecov]

Codecov Report

Merging #4914 (d67c9a3) into main (9cf800b) will increase coverage by 2.25%.
The diff coverage is 96.00%.

@@            Coverage Diff             @@
##             main    #4914      +/-   ##
==========================================
+ Coverage   74.08%   76.33%   +2.25%     
==========================================
  Files          86       86              
  Lines       13862    13925      +63     
==========================================
+ Hits        10270    10630     +360     
+ Misses       3592     3295     -297     

Impacted Files	Coverage Δ
pymc3/step_methods/hmc/base_hmc.py	91.05% <ø> (ø)
pymc3/step_methods/pgbart.py	95.80% <94.59%> (+77.67%)	⬆️
pymc3/distributions/bart.py	100.00% <100.00%> (+81.94%)	⬆️
pymc3/distributions/tree.py	100.00% <100.00%> (+69.13%)	⬆️
pymc3/sampling.py	86.00% <100.00%> (+0.32%)	⬆️
pymc3/step_methods/hmc/nuts.py	97.50% <100.00%> (ø)
pymc3/smc/smc.py	98.37% <0.00%> (-0.34%)	⬇️
pymc3/aesaraf.py	91.26% <0.00%> (ø)
pymc3/distributions/__init__.py	100.00% <0.00%> (ø)
pymc3/distributions/continuous.py	96.26% <0.00%> (+0.02%)	⬆️
... and 7 more

[aloctavodia]

This is ready to merge @ricardoV94 @twiecki?

[ricardoV94]

Looks good, but it seems some important branches in the rng_fn are not reached by the tests: https://github.com/pymc-devs/pymc3/pull/4914/files#annotation_1841363883