Fix important bug in directionalSubDiff

The directional subdifferential is assumed to be taken with respect to
a coordinate system in which v is a unit vector judging by the
implementation of subDiff in directionalconvexfunction.hh

If we do not take that into consideration, subDiff() will mix input
from two different coordinate systems and return something meaningless

As long as we only care about the sign we need not perform any
back-transformation!

src/mynonlinearity.cc

@@ -32,10 +32,9 @@ class MyNonlinearity {
       return;
     }
     double const un = u.two_norm();
-    double const ndotp = (u * v) / (un * v.two_norm());
+    double const ndotp = (u * v) / un;
+    // Our coordinate system is now such that v is a unit vector!
     if (ndotp > 0) {
-      // If we had |v| = 1, this would be
-      // <f'_pm(|u|)*u/|u|,v>
       D[1] = ndotp * func_.rightDifferential(un);
       D[0] = ndotp * func_.leftDifferential(un);
     } else {