thekla: Fix build on x86 systems without SSE support

Fixes #14709.

Same as https://github.com/Thekla/thekla_atlas/pull/11,
but adding  comments until it's merged upstream.

thirdparty/thekla_atlas/nvmath/ftoi.h

@@ -53,7 +53,10 @@ namespace nv
         return (val<0) ? ftoi_ceil_xs(val) : ftoi_floor_xs(val);
     }
 
-#if NV_CPU_X86 || NV_CPU_X86_64
+// -- GODOT start --
+//#if NV_CPU_X86 || NV_CPU_X86_64
+#if NV_USE_SSE
+// -- GODOT end --
 
     NV_FORCEINLINE int ftoi_round_sse(float f) {
         return _mm_cvt_ss2si(_mm_set_ss(f));

thirdparty/thekla_atlas/nvmath/nvmath.h

@@ -14,10 +14,12 @@
 #include <float.h>  // finite, isnan
 #endif
 
-#if NV_CPU_X86 || NV_CPU_X86_64
-    //#include <intrin.h>
-    #include <xmmintrin.h>
-#endif
+// -- GODOT start --
+//#if NV_CPU_X86 || NV_CPU_X86_64
+//    //#include <intrin.h>
+//    #include <xmmintrin.h>
+//#endif
+// -- GODOT end --
 
 
 
@@ -65,6 +67,13 @@
 #endif
 
 
+// -- GODOT start --
+#if NV_USE_SSE
+    //#include <intrin.h>
+    #include <xmmintrin.h>
+#endif
+// -- GODOT end --
+
 
 #ifndef PI
 #define PI                  float(3.1415926535897932384626433833)