tools:Move Embed/MHA/RNN/LSTM/GRU weight scale generation to ncnn2table

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

@@ -89,16 +89,16 @@ filelist_in2.txt
 ```
 **Here shape is WHC, because the order of the arguments to `ncnn::Mat`.**
 
-### 3. Quantize model
+ncnn2table can generate static weight scales without a calibration dataset for RNN,GRU,LSTM,MultiHeadAttention and Embed layers
 
 ```shell
-./ncnn2int8 mobilenet-opt.param mobilenet-opt.bin mobilenet-int8.param mobilenet-int8.bin mobilenet.table
+./ncnn2table rnn.param rnn.bin rnn.table method=kl
 ```
 
-If you don’t need static quantization, ncnn supports RNN/LSTM/GRU dynamic quantization. In this case, you can omit the table file.
+### 3. Quantize model
 
 ```shell
-./ncnn2int8 rnn-model.param rnn-model.bin rnn-model-int8.param rnn-model-int8.bin
+./ncnn2int8 mobilenet-opt.param mobilenet-opt.bin mobilenet-int8.param mobilenet-int8.bin mobilenet.table
 ```
 
 ## use ncnn int8 inference

tools/modelwriter.h

@@ -2057,6 +2057,7 @@ int ModelWriter::save(const char* parampath, const char* binpath)
             fprintf_param_value(" 4=%d", vdim)
             fprintf_param_value(" 5=%d", attn_mask)
             fprintf_param_value(" 6=%e", scale)
+            fprintf_param_value(" 7=%d", kv_cache)
             fprintf_param_value(" 18=%d", int8_scale_term)
 
             fwrite_weight_tag_data(op->q_weight_data, bp);