Add FunctionRelTol convergence test about optimize

[vladimir-ch]

... and modify DefaultSettings() to set it appropriately, e.g., DefaultSettings() could/should take FunctionInfo and set GradientAbsTol to 0 and FunctionRelTol > 0 for functions that do not evaluate the gradient.

[vladimir-ch]

Comment by @btracey copied from #34

We may need some kind of interface for FunctionRelTol like we do for Recorder. There are so many different ways to implement that kind of convergence. The most common would be that you haven’t seen an improvement of eps in Y evaluations, but there are many such flavors. I would imagine there is a big overlap between this and a future Global optimization function, where the only reasonable global convergence metrics have to do with the number of function evaluations and the last time that significant improvement has been seen. It’s worth thinking about that when doing the design.

[vladimir-ch]

Closed by d702f88 in #91