Avoid division by zero

I used nummeric_limits::max() instead of numeric_limits::infinity() here
because the header uses the former to signify an error that cannot lie
below any tolerance elsewhere already.

dune/solvers/solvers/loopsolver.cc

@@ -112,8 +112,9 @@ void ::LoopSolver<VectorType, BitVectorType>::solve()
         }
 
         // Turn the error into the relative error, if requested
-        if (this->useRelativeError_ && !std::isnan(error/oldNorm))
-            error = error / oldNorm;
+        if (this->useRelativeError_ && error != 0)
+            error = (oldNorm == 0) ? std::numeric_limits<real_type>::max()
+                                   : error / oldNorm;
 
         if (!std::isinf(convRate) && !std::isnan(convRate) && i>0)
         {