enhancements and new robust feature in line search within optimizer=5

distinguish between TolFun (optimizer termination criterion) and TolGstep/TolGstepRel, used for tuning gradient step.
So, we can keep for example TolFun=1.e-5, while allowing coarser values for TolGstep. this may help when maximizing non smooth surfaces (e.g. PKF or very large models), where numerical noise may count. By default TolGstep=TolFun as in usual historical behavior.