GHA: Add NumPy SIMD dispatch diagnostic for flaky Weber_BMC2015 (#3078)

[dweindl]

Summary

Follow-up to #3078 and #3229. The AVX-512-presence correlation reported
there was real but indirect: the actual root cause is NumPy's own runtime
SIMD dispatch for transcendental ufuncs (np.power, np.exp), used when
AMICI unscales log10/log-scaled PEtab parameters back to linear scale.
Weber_BMC2015 has all 40 of its parameters on log10 scale, so this path
is exercised heavily. NumPy's optimized kernels at different SIMD widths
(X86_V3 vs. X86_V4/AVX512_ICL/AVX512_SPR) aren't required to agree
bit-for-bit -- documented, accepted NumPy behavior -- and it silently picks
the widest tier the host CPU supports. AVX-512 presence only predicted the
outcome because it gates which tier NumPy even has the option to pick; it
is not a causal pathway through AMICI's own C++ core, SUNDIALS, SuiteSparse,
glibc, or OpenBLAS, all of which were directly tested (via causal probes on
real AVX-512 CI hardware) and refuted. Full writeup:
#3078 (comment)

This PR previously carried the exploratory causal-probe steps
(GLIBC_TUNABLES, OPENBLAS_CORETYPE=HASWELL, a full state-trajectory
dump) used to reach that conclusion. Now that the mechanism is confirmed,
those are removed, leaving only the one diagnostic that's actually useful
going forward.

What this adds

A step that prints NumPy's actual runtime SIMD dispatch state
(numpy._core._multiarray_umath.__cpu_baseline__/__cpu_dispatch__/
__cpu_features__) right after NumPy becomes available in the job. Generic
hardware diagnostics (dweindl/gha-runner-diagnostics@v1, already present,
merged in #3229) run before Python/NumPy even exist and can only show raw
CPU feature flags -- not which dispatch tier NumPy actually selected. This
shows the real variable directly, rather than requiring it to be inferred
indirectly from CPU flags as this investigation had to do.

Test plan

• Verified the exact snippet runs correctly and reports the expected
  fields locally.
• YAML validated.
• Confirm it runs cleanly in CI on both extract_subexpressions matrix
  legs.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 77.97%. Comparing base (be70933) to head (4634219).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3231      +/-   ##
==========================================
- Coverage   78.55%   77.97%   -0.58%     
==========================================
  Files         318      318              
  Lines       21106    21106              
  Branches     1488     1488              
==========================================
- Hits        16579    16458     -121     
- Misses       4519     4640     +121     
  Partials        8        8              

Flag	Coverage Δ
cpp	72.00% <ø> (+0.04%)	⬆️
cpp_python	36.56% <ø> (ø)
petab	48.29% <ø> (ø)
petab_sciml	16.28% <ø> (ø)
petab_sciml_benchmarks	14.85% <ø> (ø)
python	70.37% <ø> (+0.04%)	⬆️
sbmlsuite-jax	?

Flags with carried forward coverage won't be shown. Click here to find out more.
see 4 files with indirect coverage changes

🚀 New features to boost your workflow:

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