[IMP-dev] Minor changes to Restraint::evaluate, safe_evaluate

Restraint::evaluate() now always should return the correct answer (before it could return garbage if not called from Model::evaluate()). Internally, restraints now have a function Restraint::unprotected_evaluate() which does the actual evaluation (and is what you implement when you implement a restraint). Restraint::evaluate() makes sure the model's score states are updated before calling Restriant::unprotected_evaluate().