BART: add linear response, increase number of trees fitted per step

[aloctavodia]

This adds the option to use a linear fit to compute the values of the leaf_nodes. A new argument "response" control if the leaf_node are computed using the mean value (as previously), a linear fit or a mix. I have set "mix" by default. At in preliminary test it seems to provide a better general performance (close to just using linear). This also increase the number of trees fitted per step, increasing the computational cost, but improving diagnostics. This also includes other small fixes.

[codecov]

Codecov Report

Merging #5044 (8cd1dfe) into main (8c59b41) will increase coverage by 0.04%.
The diff coverage is 98.36%.

@@            Coverage Diff             @@
##             main    #5044      +/-   ##
==========================================
+ Coverage   78.29%   78.33%   +0.04%     
==========================================
  Files         129      129              
  Lines       24384    24427      +43     
==========================================
+ Hits        19092    19136      +44     
+ Misses       5292     5291       -1     

Impacted Files	Coverage Δ
pymc/tests/test_bart.py	100.00% <ø> (ø)
pymc/step_methods/pgbart.py	95.95% <97.77%> (+0.42%)	⬆️
pymc/distributions/bart.py	100.00% <100.00%> (ø)
pymc/distributions/tree.py	100.00% <100.00%> (ø)
pymc/__init__.py	96.00% <0.00%> (-4.00%)	⬇️
pymc/distributions/discrete.py	98.34% <0.00%> (ø)
pymc/parallel_sampling.py	87.33% <0.00%> (+1.00%)	⬆️