[Feature] Maybe only use clique in CSMC, not full tree

[Affie]

Ok, so we can reevaluate in the future:
If we remove the tree from CSMC and only keep a clique, then we can copy the channels from the tree to the CSMC instead of the full parent and child cliques.
It works by reference so the only difference will be with debugging as deep-copies are made.

Originally posted by @Affie in #675 (comment)

[dehann]

I think it would be good to drop csmc.tree if able, as long as we keep the ability to do #742

[dehann]

One reason for keeping csmc.tree is for debugging, to draw the entire tree state at each step in multithreaded system. So let's keep tree for now, but what we can do is make sure that we are only using csmc.clique, csmc.parent, csmc.children fields during calculations. We can avoid pointing back the the tree during computations.

This is not a priority now, but later perhaps we can improve the BayesTree api to have only a "shallow" version of the tree for visualization / debugging purposes, much like DFGVariableSummary.