BUG + fix: Incorrect treatment of NaNs by Geodesic::Inverse

A bug is the treatment of NaNs by Geodesic::Inverse was introduced in
version 1.25. The symptoms is that finite bogus results are returned
when a NaN is provided as one of input arguments; for example

$ echo 0 10 20 nan | GeodSolve -i
-70.06176456 -90.00000000 9987119.260

The following patch fixes this bug. This will be included in the next
release of GeographicLib.

diff --git a/src/Geodesic.cpp b/src/Geodesic.cpp
index a428ae4..d62e518 100644
--- a/src/Geodesic.cpp
+++ b/src/Geodesic.cpp
@@ -710,7 +710,8 @@ namespace GeographicLib {
         calp1 = sbet12a - cbet2 * sbet1 * Math::sq(somg12) / (1 - comg12);
       }
     }
-    if (salp1 > 0)              // Sanity check on starting guess
+    // Sanity check on starting guess.  Backwards check allows NaN through.
+    if (!(salp1 <= 0))
       SinCosNorm(salp1, calp1);
     else {
       salp1 = 1; calp1 = 0;
diff --git a/src/GeodesicExact.cpp b/src/GeodesicExact.cpp
index 47479e7..3912a24 100644
--- a/src/GeodesicExact.cpp
+++ b/src/GeodesicExact.cpp
@@ -721,7 +721,8 @@ namespace GeographicLib {
         calp1 = sbet12a - cbet2 * sbet1 * Math::sq(somg12) / (1 - comg12);
       }
     }
-    if (salp1 > 0)              // Sanity check on starting guess
+    // Sanity check on starting guess.  Backwards check allows NaN through.
+    if (!(salp1 <= 0))
       SinCosNorm(salp1, calp1);
     else {
       salp1 = 1; calp1 = 0;