patch 8.0.0614: float2nr() is not exactly right

Problem:    float2nr() is not exactly right.
Solution:   Make float2nr() more accurate.  Turn test64 into a new style test.
            (Hirohito Higashi, closes #1688)
diff --git a/src/macros.h b/src/macros.h
index 8ebc8ce..d75d190 100644
--- a/src/macros.h
+++ b/src/macros.h
@@ -364,6 +364,9 @@
 #  if !defined(NAN)
 #   define NAN (INFINITY-INFINITY)
 #  endif
+#  if !defined(DBL_EPSILON)
+#   define DBL_EPSILON 2.2204460492503131e-16
+#  endif
 # endif
 #endif