multiheadattention int8 quantization

[nihui]

• arch fallback
• quantization tool
• coverage++
• code clean
• op doc

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 98.03279% with 6 lines in your changes missing coverage. Please review.

Project coverage is 94.68%. Comparing base (1c7af00) to head (85ddd8a).
Report is 3 commits behind head on master.

Files with missing lines	Patch %	Lines
src/layer/multiheadattention.cpp	97.35%	6 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #5733      +/-   ##
==========================================
- Coverage   95.15%   94.68%   -0.47%     
==========================================
  Files         793      553     -240     
  Lines      270315   200255   -70060     
==========================================
- Hits       257218   189618   -67600     
+ Misses      13097    10637    -2460     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[nihui]

mha * 12 (ms)	fp32	fp16	bf16 (disabled atm)	int8/fp32	int8/fp16	int8/bf16 (disabled atm)
RES (kb)	51180	27816	27744	15380	15744	15548
qcom855plus big core 1t	34.68	18.59	46.30	11.69	11.70	11.94
qcom855plus small core 1t	188.14	104.37	176.96	62.78	59.10	61.01
mtk9000 big core 1t	60.12	30.83	69.63	13.00	12.77	12.89
mtk9000 small core 1t	153.16	103.58	181.27	49.55	45.01	50.39
imx7d 1t	863.34		1987.19	452.76		445.55

thus, armv7 gemm needs inline assembly optimization