Small runtime improvement for hybrid

[dellaert]

[Targeting branch from #1843, develop after that's merged]

Using instruments I was profiling testHybridEstimation, and I discovered a small gain (a few %) by making the DecisionTreeFactor::combine a bit more efficient.

However, @varunagrawal, I also noticed that augment in HGF is one of the heaviest hitters - probably 10-20%. This is not surprising, as this where the "hiding" happens, which is expensive and not needed. The negLogConstants can just be added during elimination,and we probably gain in the QR calls as well. The hiding has no benefit except that we would not have to rewrite a tiny part of the math :-/

[dellaert]

PS those test write nice pdf files if graphviz is installed. I had to fix something in DecisionTree, though:

[varunagrawal]

Interesting. I was under the impression that "hiding" would be beneficial since then it's all Matrix-Vector code which Eigen can optimize really well. Guess not...