unify the placement of transpose op (#895)

benchmark/alexnet/1_machine_1_gpu.placement

 placement_group {
   op_set {
     op_name: "decode"
-    op_name: "transpose"
   }
   parallel_conf {
     policy: kDataParallel
@@ -11,6 +10,7 @@ placement_group {
 
 placement_group {
   op_set {
+    op_name: "transpose"
     op_name: "conv1"
     op_name: "pool1"
     op_name: "conv2"