objective occassionally decreases

[stephens999]

Sometimes the objective function (variational lower bound) decreases in an iteration.
This should not happen in principle, but it does - usually it is just a very small amount.

Possible reasons include: choice of grid
is not the same from iteration to iteration (when using ebnm_ash); the optimization is
converging to a local optima (when using ebnm_pn). If these reasons are correct then
possibly we could avoid the problem by using the current g to initialize the solution of the ebmn problem.