BART: clamp first particle to old full tree

[aloctavodia]

We no longer keep track of the full history of a particle/tree, we just keep the old tree (so there is still a chance to do not accept a move/new tree). The effect on the function approximation is negligible, but instead the sampler is faster and also the code becomes simpler.

[codecov]

Codecov Report

Merging #5011 (4888801) into main (641b278) will increase coverage by 0.48%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main    #5011      +/-   ##
==========================================
+ Coverage   77.82%   78.30%   +0.48%     
==========================================
  Files         128      129       +1     
  Lines       24380    24384       +4     
==========================================
+ Hits        18973    19095     +122     
+ Misses       5407     5289     -118     

Impacted Files	Coverage Δ
pymc/step_methods/pgbart.py	95.52% <100.00%> (-0.28%)	⬇️
pymc/step_methods/hmc/integration.py	78.84% <0.00%> (-11.54%)	⬇️
pymc/backends/report.py	87.41% <0.00%> (-4.20%)	⬇️
pymc/step_methods/hmc/nuts.py	95.00% <0.00%> (-2.50%)	⬇️
pymc/tests/test_step.py	48.63% <0.00%> (-0.96%)	⬇️
pymc/tests/test_moment.py	100.00% <0.00%> (ø)
pymc/tests/test_sampling.py	97.29% <0.00%> (+0.23%)	⬆️
pymc/tests/test_distributions.py	96.34% <0.00%> (+0.33%)	⬆️
pymc/sampling.py	87.42% <0.00%> (+0.40%)	⬆️
pymc/tests/test_model.py	99.11% <0.00%> (+0.88%)	⬆️
... and 8 more

[michaelosthege]

I have just a few nitpicks related to code style.
Since I'm not familiar with the sampler, I can't tell if you're breaking it, so take my approval with a grain of salt ;)