optimize compilation time of argmin/argmax op (#29595)

* Using VisitDataTypeTiny and put CastOP after ReduceOP, test=develop

* remove changes of reduce_op.h, test=develop

paddle/fluid/operators/arg_min_max_op_base.cu.h

@@ -175,12 +175,13 @@ class ArgMinMaxOpCUDAKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto& dtype = ctx.Attr<int>("dtype");
     if (dtype < 0) {
-      framework::VisitDataType(static_cast<framework::proto::VarType::Type>(
-                                   framework::proto::VarType::INT64),
-                               VisitDataCudaArgMinMaxFunctor<T, Reducer>(ctx));
+      framework::VisitDataTypeTiny(
+          static_cast<framework::proto::VarType::Type>(
+              framework::proto::VarType::INT64),
+          VisitDataCudaArgMinMaxFunctor<T, Reducer>(ctx));
       return;
     }
-    framework::VisitDataType(
+    framework::VisitDataTypeTiny(
         static_cast<framework::proto::VarType::Type>(dtype),
         VisitDataCudaArgMinMaxFunctor<T, Reducer>(ctx));
   }

paddle/fluid/operators/arg_min_max_op_base.h

@@ -128,13 +128,13 @@ class ArgMinMaxKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const override {
     auto& dtype = ctx.Attr<int>("dtype");
     if (dtype < 0) {
-      framework::VisitDataType(
+      framework::VisitDataTypeTiny(
           static_cast<framework::proto::VarType::Type>(
               framework::proto::VarType::INT64),
           VisitDataArgMinMaxFunctor<DeviceContext, T, EnumArgMinMaxValue>(ctx));
       return;
     }
-    framework::VisitDataType(
+    framework::VisitDataTypeTiny(
         static_cast<framework::proto::VarType::Type>(dtype),
         VisitDataArgMinMaxFunctor<DeviceContext, T, EnumArgMinMaxValue>(ctx));
   }