Fix out-of-bounds access in is_hermsym for stored zeros

[KristofferC]

When old sins come back to haunt you.... (I am responsible for this bug 11 years ago)... I made Claude do some archaeology:

---

History of the is_hermsym out-of-bounds bug

The specific crashing loop is from JuliaLang/julia#15504, but the bug lineage starts at JuliaLang/julia#11371 — a version of this out-of-bounds hole has existed continuously since May 2015. Here is what an empirical run of each historical version of the algorithm shows:

Case	#11371 (2015)	#15504 (2016)	#16526 (2016)	truth
Today's crasher (stored zero, partner column runs out)	wrong answer: true	BoundsError	BoundsError	false
Exhausted partner column, no stored zeros	BoundsError	BoundsError	false ✓	false
#15504 matrix	wrong answer: false	true ✓	true ✓	true
Issue #16521 matrix	BoundsError	BoundsError	false ✓	false

The history reads like a game of whack-a-mole:

1. JuliaLang/julia#11371 ("Improve performance of issym and ishermitian", May 2015) introduced is_hermsym with an unguarded rowval[tracker[row]] read — already out-of-bounds-capable (reported as JuliaLang/julia#16521), and it returned a silently wrong true on today's test matrix.
2. JuliaLang/julia#15504 ("fix bug when zeros are stored", March 2016) fixed the stored-zeros wrong answers by adding the catch-up loop — and that loop's unguarded offset += 1; row2 = rowval[offset] is exactly the line that throws in today's crasher. So the precise bug fixed here was born there.
3. JuliaLang/julia#16526 ("fix bug in sparse is_hermsym", May 2016, fixing JuliaLang/julia#16521) added the offset > length(rowval) guard — but only before the first read, and with the wrong bound (end of rowval instead of end of the column). It fixed the no-stored-zeros case and left the stored-zeros hole open for another decade.
4. #661 (2023) re-added @inbounds, quietly upgrading the remaining BoundsError to undefined behavior.

So if you want a single answer: JuliaLang/julia#15504 wrote the line that crashes, but it was replacing a JuliaLang/julia#11371 line that gave a wrong answer on the same input. This fix is the third patch to the same ten-line region, all chasing the original 2015 tracker design.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.35%. Comparing base (c50786b) to head (e39e2ee).

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #741   +/-   ##
=======================================
  Coverage   84.34%   84.35%           
=======================================
  Files          13       13           
  Lines        9373     9375    +2     
=======================================
+ Hits         7906     7908    +2     
  Misses       1467     1467           

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[dkarrasch]

Apparently, we need to backport to all the way to v0.4!

[KristofferC]

I ran out of labels!