BART: improve sampling

[aloctavodia]

This has essentially the same effect as the changes proposed in #5223 but avoiding the horrible heuristic of resetting the sampler. Instead, one the particle is copied from the previous tree and the leaf values are resampled taking into account the current sum of trees. This PR also increases the number of particles but decreases the number of tress fitted per step. The overall effect is a slower sampler that we currently have but much faster than the change proposed in #5223. Still hold the observation that proposing better trees should lead to decrease the number of necessary particles while keeping the accuracy high and the sampler running smoothly. That is something I will explore in a future PR.

[codecov]

Codecov Report

Merging #5229 (f023e32) into main (9d4691c) will decrease coverage by 0.04%.
The diff coverage is 100.00%.

❗ Current head f023e32 differs from pull request most recent head 45d5afb. Consider uploading reports for the commit 45d5afb to get more accurate results

@@            Coverage Diff             @@
##             main    #5229      +/-   ##
==========================================
- Coverage   78.98%   78.93%   -0.05%     
==========================================
  Files          88       88              
  Lines       14273    14232      -41     
==========================================
- Hits        11273    11234      -39     
+ Misses       3000     2998       -2     

Impacted Files	Coverage Δ
pymc/bart/bart.py	95.91% <ø> (ø)
pymc/bart/pgbart.py	95.60% <100.00%> (+0.45%)	⬆️
pymc/bart/tree.py	100.00% <100.00%> (ø)

[junpenglao]

Dont forget the release note :)