[apullin]

Fixes #6224.

Each thread of the threaded SYRK/HERK driver packs one K block of op(A) for its whole row range behind sb in its work area.
For the widest thread that is about GEMM_Q * N / sqrt(nthreads) elements.
Nothing limited it to the space left in the BUFFER_SIZE work area, so for large N it wrote past the end (in #6224: into C and its malloc header).
Threaded ?POTRF uses the same driver and was affected too.

The threshold scales with BUFFER_SIZE: 32 MiB in scipy-openblas (BUFFERSIZE=20), 128 MiB by default.
A default x86_64 build overflows too, at larger N (Haswell kernels, 2 threads: from N ≈ 92,000).

The fix shortens the K block until the widest thread's panels fit behind sb.
All threads use the same blocking, since they read each other's panels.
Where the panels already fit, nothing changes.

Tests:

• utest/test_syrk_thread.c checks threaded ?SYRK/?HERK (all variants) and ?POTRF against an exact integer reference; with a work area of 16 MiB or less (32-bit x86) it also runs a ZHERK that overflowed.
• utest/test_syrk_thread_workarea.c calls the threaded driver with a work area the test owns, at the smallest N that overflowed (a few seconds natively; needs the static library; skipped above OPENBLAS_UTEST_TIME_LIMIT, default 10 s, e.g. under qemu). It uses driver internals, so it is a separate commit you can drop.

Tested on develop c2fc6cf, before → after:

• the issue's program (BUFFERSIZE=20 INTERFACE64=1, SkylakeX kernels under Intel SDE, 12 threads): SIGSEGV → all elements correct
• default x86_64 build, DSYRK N=93000, 2 threads: SIGSEGV → all elements correct
• 32-bit x86, 2 threads: ZHERK N=5500 heap corruption → correct; ZPOTRF N=6000 crash → correct
• work_area_bound: fails → passes on x86_64 and 32-bit x86, and with the time limit lifted on arm64, POWER, riscv64 and LoongArch (qemu)
• no other test result changes: utest on 10 architectures; BLAS, CBLAS and LAPACK tests on x86_64; Make and CMake (static/shared), OpenMP, USE_THREAD=0, ILP64

Where SYRK is already broken (SVE wider than the kernels assume, #6100; RISCV64_ZVL256B at VLEN ≥ 512; 32-bit Penryn/Dunnington), the new exactness tests fail before and after, like existing tests.

[apullin]

The two msys2 CLANG64 failures are not from this change.
The new utest is the first to run SSYRK/SGEMM with K > 341 on clang/Windows, where the AVX512 fallback kernel overflows its stack buffer: the SGEMM twin of #6232.
Fix: #6236. With it and #6234, the CLANG64 jobs pass (checked on my fork).
I'll update this branch once both are in develop.