Use selfAdjoingView<Upper> in InvertPSDMatrix.

This restores some sensitive numerical tests.

Change-Id: I7e6fc0236e810b1c112af75067957740c40c8f36

internal/ceres/invert_psd_matrix.h

diff --git a/internal/ceres/invert_psd_matrix.h b/internal/ceres/invert_psd_matrix.h
index 60f6a57..bbf52fe 100644
--- a/internal/ceres/invert_psd_matrix.h
+++ b/internal/ceres/invert_psd_matrix.h
@@ -61,9 +61,10 @@
   // https://eigen.tuxfamily.org/dox/group__TutorialLinearAlgebra.html#title3
   if (assume_full_rank) {
     if (kSize > 0 && kSize < 5) {
-        return m.inverse();
+      return m.inverse();
     }
-    return m.llt().solve(MType::Identity(size, size));
+    return m.template selfadjointView<Eigen::Upper>().llt().solve(
+        MType::Identity(size, size));
   }
 
   Eigen::JacobiSVD<MType> svd(m, Eigen::ComputeThinU | Eigen::ComputeThinV);