Bart: Refactor splitting variables and predictions

[aloctavodia]

This add a new predict method to compute out of sample predictions evaluated at new X values. This also refactors how splitting variables are generated, this simplifies the previous code and provides a small performance improvement.

[codecov]

Codecov Report

Merging #4310 (8e94ed7) into master (0431292) will increase coverage by 0.57%.
The diff coverage is 60.56%.

@@            Coverage Diff             @@
##           master    #4310      +/-   ##
==========================================
+ Coverage   87.56%   88.14%   +0.57%     
==========================================
  Files          88       87       -1     
  Lines       14270    14361      +91     
==========================================
+ Hits        12496    12658     +162     
+ Misses       1774     1703      -71     

Impacted Files	Coverage Δ
pymc3/distributions/tree.py	87.65% <14.28%> (-0.96%)	⬇️
pymc3/distributions/bart.py	70.12% <58.49%> (-10.68%)	⬇️
pymc3/step_methods/pgbart.py	98.10% <100.00%> (+0.11%)	⬆️
pymc3/distributions/discrete.py	94.25% <0.00%> (-2.54%)	⬇️
pymc3/parallel_sampling.py	85.84% <0.00%> (-0.95%)	⬇️
pymc3/data.py	80.26% <0.00%> (-0.67%)	⬇️
pymc3/distributions/posterior_predictive.py	89.17% <0.00%> (-0.13%)	⬇️
pymc3/variational/approximations.py	80.12% <0.00%> (-0.07%)	⬇️
pymc3/variational/flows.py	93.58% <0.00%> (-0.02%)	⬇️
pymc3/step_methods/mlda.py	95.50% <0.00%> (-0.02%)	⬇️
... and 24 more

[michaelosthege]

I can't really tell what the changes are about. Please add some description of intent/purpose for this PR.
Also @aloctavodia please propose a line to be added to the README over in #4424 that mentions this PR and informs about the API change.