fix ncnn2int8 top_blob int8 scales saving for dequant only style (#6425)

Author: nihui

docs/how-to-use-and-FAQ/quantized-int8-inference.md

@@ -8,6 +8,8 @@ Example with mobilenet, just need three steps.
 
 ### 1. Optimize model
 
+NOTE: **If your model is converted via pnnx, skip this step.**
+
 ```shell
 ./ncnnoptimize mobilenet.param mobilenet.bin mobilenet-opt.param mobilenet-opt.bin 0
 ```

tools/modelwriter.h

@@ -900,7 +900,10 @@ int ModelWriter::save(const char* parampath, const char* binpath)
                 {
                     fwrite_weight_data(op->weight_data_int8_scales, bp, 90, 100);
                     fwrite_weight_data(op->bottom_blob_int8_scales, bp, 0.001, 1);
-                    fwrite_weight_data(op->top_blob_int8_scales, bp, 0.001, 1);
+                    if (op->int8_scale_term > 100)
+                    {
+                        fwrite_weight_data(op->top_blob_int8_scales, bp, 0.001, 1);
+                    }
                 }
 #endif // NCNN_INT8
             }
@@ -1073,7 +1076,10 @@ int ModelWriter::save(const char* parampath, const char* binpath)
                 {
                     fwrite_weight_data(op->weight_data_int8_scales, bp, 90, 100);
                     fwrite_weight_data(op->bottom_blob_int8_scales, bp, 0.001, 1);
-                    fwrite_weight_data(op->top_blob_int8_scales, bp, 0.001, 1);
+                    if (op->int8_scale_term > 100)
+                    {
+                        fwrite_weight_data(op->top_blob_int8_scales, bp, 0.001, 1);
+                    }
                 }
 #endif // NCNN_INT8
             }