MOSES: Missing multivariate algo

[linas]

Moshe's thesis describes bayesian optimization algorithms/bayesian distribution estimation; however, these seem to have somehow not been kept alive/ported over to the current moses code base. These should be revived/re-implemented. So, for example: in learning/moses/eda/local_structure.h, we see that univariate::operator() is a no-op: there is no structure learning whatsoever. Thus, dependencies between different knobs/variables is never accounted for, making learning more-or-less impossible for "deceptive" problems.

As a demo of the issue, I wrote learning/moses/example-progs/trap-bits.cc to illustrate boa learning in moses, before discovering that it's not currently implemented :-(