Fix sign guard on the cliq liquid extrapolation in thermr

[ramic-k]

Fixes #399.

What

The small-α extrapolation in sig(),

s = sab(1,1) + log(alpha(1)/a)/2 - cliq*b**2/a

assumes cliq >= 0, which holds when S(α₁,β) decreases from β₁ to β₂.
The activation guard tests decay along the α axis instead
(sab(1,1) > sab(2,1)). A scattering law whose β=0 row decays in α
while S(α₁,β) rises from β=0 to β₂ — e.g. a coherent one-phonon law
for a crystalline powder, where the acoustic intensity at small Q sits
in the first nonzero β bin — makes cliq negative, so the
extrapolation grows like exp(+|cliq|·b²/a) with a clamped at 1e-6
and s clamped only from below. The result is unphysical thermal
inelastic cross sections (up to ~1e150+ barns with the reproducer in
the linked issue; ~1e91 b with the real graphite evaluation we first
hit this on) or runaway adaptive refinement.

This PR requires decay along both axes before activating cliq, at
both computation sites:

if (sab(1,1).gt.sab(2,1).and.sab(1,1).gt.sab(1,2))&
  cliq=(sab(1,1)-sab(1,2))*alpha(1)/beta(2)**2

Why existing results are unaffected

For liquid evaluations — the intended target of the extrapolation —
both decays hold (checked explicitly for ENDF/B-VIII.0 H-in-H₂O:
S(α₁,β₁)=6.17e5 > S(α₁,β₂)=6.03e3), so cliq keeps its current
value. Materials that never activated the guard keep cliq=0. Only the
pathological negative-cliq case changes, from garbage to the SCT/
tabulated treatment of the same region.

On macOS/arm64 the repo test suite gives identical results with and
without this change (same pass/fail sets; the 14 failures present on
this platform are present on unpatched develop as well).