add output defs for send_ue_recv kernel (#51522)

paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc

@@ -89,7 +89,6 @@ static std::set<std::string> OpsNeedSetOutputDtypeWhenRegisterPhiKernel = {
     "rnn",
     "search_sort",
     "select",
-    "send_ue_recv",
     "sync_batch_norm_grad",
     "unique",
     "unique_consecutive_flattened_tensor",

paddle/phi/kernels/cpu/send_ue_recv_kernel.cc

@@ -290,4 +290,6 @@ PD_REGISTER_KERNEL(send_ue_recv,
                    float,
                    double,
                    int,
-                   int64_t) {}
+                   int64_t) {
+  kernel->OutputAt(1).SetDataType(phi::DataType::INT32);
+}

paddle/phi/kernels/gpu/send_ue_recv_kernel.cu

@@ -332,4 +332,6 @@ PD_REGISTER_KERNEL(send_ue_recv,
                    double,
                    int,
                    int64_t,
-                   phi::dtype::float16) {}
+                   phi::dtype::float16) {
+  kernel->OutputAt(1).SetDataType(phi::DataType::INT32);
+}