support layernorm int8

[tpoisonooo]

基于 PR #4096 接着写。

• 根据 FQViT 论文实现了 int8 layernorm
  • imagenet1k 数据集，只量化 LN 精度 - 0.3%； 量化 LN+Conv+Gemm+mha + bias correlation 精度 -0.9%
  • ncnn2int8 开发了 layernorm requant 功能， 匹配 layernorm_int --> ( mha_int8 | conv_int8 | gemm_int8 ) 模式
  • 更新了 operators.md 文档
  • 更新了对应的 testcase
• example 增加 vision_transformer_int8，模型下载路径在注释里了
• 多个 layer 都写了一次的 float2int8 放到了公共的 mathfunction.h 里

失败的尝试，已经 revert 掉了：

1. 尝试 int8 binaryop ，精度下降严重
2. 由于 innerproduct_int8 没实现 requant，导致图优化做得很烂、int8 binaryop 没加速效果

[codecov-commenter]

Codecov Report

Merging #4132 (81a1149) into master (b4ba207) will decrease coverage by 1.91%.
The diff coverage is 94.67%.

@@             Coverage Diff             @@
##           master    #4132       +/-   ##
===========================================
- Coverage   94.42%   92.51%    -1.92%     
===========================================
  Files         747      374      -373     
  Lines      178769    66122   -112647     
===========================================
- Hits       168811    61173   -107638     
+ Misses       9958     4949     -5009     

Impacted Files	Coverage Δ
src/layer/convolution.cpp	94.24% <ø> (-0.22%)	⬇️
src/layer/layernorm.cpp	93.08% <85.52%> (-5.13%)	⬇️
src/layer/multiheadattention.cpp	98.53% <98.77%> (+5.30%)	⬆️
src/layer/mathfun.h	100.00% <100.00%> (ø)
src/layer/arm/convolution_winograd_transform.h	0.00% <0.00%> (-100.00%)	⬇️
...c/layer/arm/convolution_winograd_transform_bf16s.h	0.00% <0.00%> (-98.31%)	⬇️
src/layer/arm/flatten_arm.cpp	35.74% <0.00%> (-63.46%)	⬇️
src/layer/arm/packing_arm.cpp	64.53% <0.00%> (-31.76%)	⬇️
src/layer/arm/slice_arm.cpp	80.32% <0.00%> (-17.15%)	⬇️
src/layer/arm/concat_arm.cpp	80.32% <0.00%> (-16.25%)	⬇️
... and 491 more

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[lgtm-com]

This pull request introduces 1 alert when merging 9277a1f into acbaaa6 - view on LGTM.com

new alerts:

• 1 for FIXME comment