fix softmax arm fp16s sum error, fix #5340

[nihui]

No description provided.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 94.48%. Comparing base (84256b1) to head (436ebaf).
Report is 3 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #5393      +/-   ##
==========================================
- Coverage   94.63%   94.48%   -0.16%     
==========================================
  Files         778      778              
  Lines      243359   243358       -1     
==========================================
- Hits       230302   229925     -377     
- Misses      13057    13433     +376     

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

[glenn-jocher]

@nihui would it be possible to publish new NCNN and/or PNNX versions with this fix? This may help improve Ultralytics YOLO export validation accuracies with NCNN which are trailing other formats:

https://github.com/ultralytics/ultralytics/actions/runs/8584738579/job/23525336246

[nihui]

@nihui would it be possible to publish new NCNN and/or PNNX versions with this fix? This may help improve Ultralytics YOLO export validation accuracies with NCNN which are trailing other formats:

https://github.com/ultralytics/ultralytics/actions/runs/8584738579/job/23525336246

New versions of ncnn and pnnx will be released this week
Currently trying to fix more bugs to ensure stability and compatibility, across gpu risc-v mips etc