Fedora 45 in CI: groff 1.24, no col, OpenBLAS's batch crash with DYNAMIC_ARCH #330

Since 2026-11-03, fedora:latest is Fedora 45, and the man and flexiblas jobs fail. Main would fail the same way; seen on #328. There are three causes:

• groff 1.24.1 re-reads the hyphenation mode from the HY register at every paragraph, so the pages' .nh no longer holds. test_man_render.sh finds a URL hyphenated across a line (blas-microbench‐mark). Fix: .nr HY 0 after .TH.
• col is missing. It is no longer pulled in by man-db under install_weak_deps=False, so man fails in test_man_install.sh. Fix: install util-linux.
• OpenBLAS 0.3.34 segfaults on bmb_dgemm --batch-mode strided -m 8 (test_man_content.sh, exit 139). OpenBLAS 0.3.28 to 0.3.34, built with DYNAMIC_ARCH (every distribution builds it that way), calls the small-matrix kernel's offset instead of its address in a batch: the grouped batch from 0.3.28 (Fix: Small matrix gemm_batch with DYNAMIC_ARCH OpenMathLib/OpenBLAS#5980), the strided one from 0.3.31 (Fix complex scalar pointers in strided batched GEMM OpenMathLib/OpenBLAS#6086). It is fixed for 0.3.35. I reproduced it with a 0.3.34 DYNAMIC_ARCH build on every core type, and not with a build for one target. Decided: the benchmark recognises such an OpenBLAS and refuses a batch of small matrices before calling it: exit 77, saying why and naming the upstream issue. The fix also adds a Known issues entry, the man page, and tests.

Also: OpenMathLib/OpenBLAS#6214 (complex strided batch strides), which bmb_known_wrong takes as having no upper bound, was fixed for 0.3.35 by OpenMathLib/OpenBLAS#6215. Its bound goes in, once verified on a build of the fix.

Blocks the 1.10.0 release (#325).