OK, so now we know there are tiny differences between distances

computed by ANN's distance function and geogram distance function, may
come from some automatic FP optimizations triggered by default by Mac
clang. So we have a solution that works with a tiny tolerance but I do
not like it. Here I'm just trying to recompute geogram's distances using
ANN's distance functions before comparing. We shall see...

src/tests/test_nn_search/main.cpp

@@ -152,11 +152,22 @@ int main(int argc, char** argv) {
 
             bool has_mismatch = false;
 
-            for(index_t j = 0; j < nb_neigh; ++j) {
+            // Recompute distance between i and nearest neighbors computed by geogram using
+            // ANN's distance function, because there can be tiny differences if the compiler
+            // does not optimize both functions the same way (happens on Mac/M1)
+            for(index_t j=0; j < nb_neigh; ++j) {
+                index_t nn = neigh2[j];
+                sq_dist2[j] = annDist(
+                    M.vertices.dimension(), M.vertices.point_ptr(i), M.vertices.point_ptr(nn)
+                );
+            }
+
+            for(index_t j=0; j < nb_neigh; ++j) {
                 // Added tolerance: on Mac/M1 we got tiny differences,
                 // I think it is doing auto FMA here and there, to be
                 // checked.
-                if(::fabs(sq_dist1[j] - sq_dist2[j]) > 1e-6) {
+                if(sq_dist1[j] != sq_dist2[j]) {
+                // if(::fabs(sq_dist1[j] - sq_dist2[j]) > 1e-6) {
                     has_mismatch = true;
                     match = false;
                     Logger::err("Mismatch") << i << "[" << j << "]"