Allow training separate tree structures if training multiple trees

[velochy]

This adds a separate_trees flag that allows training non-shared tree structures if training multiple trees.
This is considerably slower (roughly multiplies the time taken by the number of trees being trained) but can result in drastic decreases in variability of results.

Attached is an example of results of BART with shared structure (joint), BART with separate trees and GLM all trained on our polling data. As you can see, in this case, the separation really matters a lot, so hopefully this serves as a motivating example:

[velochy]

As @aloctavodia on vacation and this did greatly improve a fit with categorical variables, maybe you could have a look her as well @juanitorduz ?

[aloctavodia]

A rebase is needed.

[velochy]

Rebase done and all checks pass. Should be good to merge once you've reviewed what it does.

[aloctavodia]

Thanks @velochy!