CI: add -Werror strict lane and an auto-discovered regression-test harness - #350

Two pieces of CI hardening (#315):

* WERROR opt-in (Make.inc) and a dedicated `strict-werror` CI job that
  builds with -Werror under both gcc and clang on a pinned ubuntu-24.04.
  It is off by default and pinned so a future compiler upgrade surfaces
  as one isolated failure rather than breaking the whole matrix. Fixes
  the warnings this surfaces in openlibm's own code:
    - irint(): mark the header-defined helper __unused (it is only used
      by e_rem_pio2*.c, so it is unused in every other TU).
    - lgamma_r/lgammal_r: zero-init nadj, which is written and read under
      the same hx<0 guard (-Wmaybe-uninitialized false positive).
  Clang-only -Wimplicit-int-float-conversion (intentional lround
  boundary trick) and -Wignored-attributes (pre-existing extern-inline
  fenv redeclarations) are documented and suppressed for the lane.

* Regression harness: test/regression/*.c are self-contained programs
  (exit 0=pass, 77=skip, else=fail) auto-discovered and run by
  `make test` from the repo root, so later fixes can land a regression
  test by dropping in one file with no Makefile changes. The previously
  orphaned test-211.c (never built or run) is moved in as the first one.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

The strict lane surfaced more on the CI compilers than on older local
ones (verified against gcc 13 / clang 18):

* gcc: zero-init nadj in the float lgammaf_r too (same guarded
  false-positive as the double/long-double versions).
* clang 18 renamed the int->float conversion diagnostic to
  -Wimplicit-const-int-float-conversion; suppress that spelling as well.
* clang: suppress -Wdeprecated-non-prototype (C23) for the lane; the K&R
  definitions in w_cabs*.c and bsdsrc/b_*.c are left for a focused
  prototype-conversion follow-up.

Cross CI: build the regression programs with the cross toolchain in the
"Build tests" step and pass the toolchain vars to the qemu run, so the
foreign-arch regression binaries are produced by $TRIPLE-gcc instead of
the host compiler. Verified locally by cross-building for riscv64 and
running the regression test under qemu.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>