Changed the stopping strategy for the optimization inside the

train_probabilistic_decision_function() because it could potentially
never stop due to large rounding error when working with very large
datasets.

--HG--
extra : convert_revision : svn%3Afdd8eb12-d10e-0410-9acb-85c331704f74/trunk%403793

dlib/svm/svm.h

@@ -822,7 +822,7 @@ namespace dlib
         matrix<double,2,1> val;
         val = 0;
         find_min(newton_search_strategy(obj_hessian),
-                 gradient_norm_stop_strategy(),
+                 objective_delta_stop_strategy(),
                  obj,
                  obj_der,
                  val,