[flang][cuda] Fix descriptor sync in data transfer (#129333)

The destination descriptor on the device needs to be sync with the
destination descriptor on the host, not the src one.

flang-rt/lib/cuda/memory.cpp

@@ -147,7 +147,7 @@ void RTDECL(CUFDataTransferGlobalDescDesc)(Descriptor *dstDesc,
     void *deviceAddr{
         RTNAME(CUFGetDeviceAddress)((void *)dstDesc, sourceFile, sourceLine)};
     RTNAME(CUFDescriptorSync)
-    ((Descriptor *)deviceAddr, srcDesc, sourceFile, sourceLine);
+    ((Descriptor *)deviceAddr, dstDesc, sourceFile, sourceLine);
   }
 }
 }