Make the valgrind CI job actually check the test binaries

[ACSimon33]

Summary

The memory-check CI job has been reporting a clean bill of health without ever
looking at a LAPACK binary. Every test is registered as a cmake -P runtest.cmake
wrapper invocation, so ctest -T memcheck ran valgrind on cmake, and the real
test executable — started by execute_process() inside the wrapper — was never
instrumented at all.

This PR fixes the instrumentation, adds an explicit LAPACK_MEMORY_CHECK option to
configure the test suite for it, and fixes the first three defects the working
memory check reported.

The fix

• Under LAPACK_MEMORY_CHECK on Unix the command is sh -c 'exec <binary> …'.
  exec replaces the shell instead of forking it, so exactly one process ever
  writes to the log file, and that process is the test binary.
• Everywhere else it keeps the existing runtest.cmake wrapper unchanged, so
  Windows, multi-config generators and failing-test output all behave as before.

The three call sites — BLAS/TESTING, CBLAS/testing and TESTING — now go
through this one helper instead of each spelling out the wrapper invocation, which
is also how the four near-duplicate add_test() blocks collapse.

New option: LAPACK_MEMORY_CHECK

• hard-requires valgrind (REQUIRED) instead of quietly doing nothing;
• forces LAPACK_TESTING_USE_PYTHON=OFF, because the Python summary is a ctest test of its own and would otherwise have valgrind profile the interpreter rather than LAPACK;
• switches the test commands to the exec form described above;
• adds --trace-children=yes, and narrows --show-leak-kinds from all to definite,indirect to keep the still-reachable noise out of the report.

The CI job in .github/workflows/special.yml now passes
-D LAPACK_MEMORY_CHECK:BOOL=ON in place of the old
-D LAPACK_TESTING_USE_PYTHON:BOOL=OFF.

Defects the working memory check found

cblas_sgemv read an uninitialized TA and called Fortran with it.
The column-major branch called cblas_xerbla() for an illegal TransA and then
fell straight through to F77_sgemv(F77_TA, …) with TA never assigned — because
return was missing. cblas_xerbla does not abort under the test harness, so this
was a live uninitialized read followed by a Fortran call with a garbage TRANS
character. The row-major branch of the same routine, and both branches of dgemv,
cgemv and zgemv, already had the return; only cblas_sgemv was missing it.

IWORK was leaked by all four LIN test drivers.
{s,d,c,z}chkaa.F allocate IWORK(34*NMAX) and deallocate E, S, WORK and
RWORK on exit but never IWORK.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 69.01%. Comparing base (dd51569) to head (3989868).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1368   +/-   ##
=======================================
  Coverage   69.01%   69.01%           
=======================================
  Files        6122     6122           
  Lines      486123   486124    +1     
  Branches    23286    23286           
=======================================
+ Hits       335514   335515    +1     
  Misses     150420   150420           
  Partials      189      189           

Components	Coverage Δ
BLAS	97.94% <ø> (ø)
CBLAS	96.98% <100.00%> (+<0.01%)	⬆️
LAPACK	82.37% <ø> (ø)
LAPACKE	0.10% <ø> (ø)
TMGLIB	55.69% <ø> (ø)
BLAS testing	88.33% <ø> (ø)
CBLAS testing	89.63% <ø> (ø)
LAPACK testing	82.35% <ø> (ø)
LAPACKE testing	∅ <ø> (∅)

Files with missing lines	Coverage Δ
CBLAS/src/cblas_sgemv.c	100.00% <100.00%> (ø)
TESTING/LIN/cchkaa.F	74.32% <ø> (ø)
TESTING/LIN/dchkaa.F	73.91% <ø> (ø)
TESTING/LIN/schkaa.F	73.91% <ø> (ø)
TESTING/LIN/zchkaa.F	73.02% <ø> (ø)

---

Continue to review full report in Codecov by Harness.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update dd51569...3989868. Read the comment docs.