Use binary search for the S(alpha,beta) interval lookup in thermr

[ramic-k]

What

sig() located the bracketing α and β intervals with linear scans on
every evaluation:

do i=1,nb1
   ib=i
   if (bbb.lt.beta(i+1)) exit
enddo

For modern fine-grid tabulations these scans dominate THERMR runtime:
profiling a coherent-crystal evaluation with a 399 α × 5001 β law put
~70% of all samples inside the two scan loops. This PR replaces both
with binary searches that select exactly the same interval (the
first i with value < grid(i+1), saturating at n-1), making the
lookup O(log n) per evaluation.

Effect

• Output is bit-for-bit identical — the selected (ia, ib) pair is
  the same for every query, including the equality and saturation edge
  cases, so all downstream arithmetic is unchanged. Verified by byte
  comparison of full RECONR/BROADR/THERMR/ACER output chains on three
  different TSL tapes (incoherent and coherent laws, 399×400 to
  399×5001 grids).
• On the small tabulations in the repo test suite the change is
  performance-neutral (and the suite results are identical with and
  without it; same pass/fail sets on macOS/arm64).
• On the 399×5001 research tape above, THERMR wall time dropped from
  39 to 23 minutes at -O0 (the scans shrink from ~70% of samples to a
  negligible share; remaining time is interpolation and quadrature).
  Large-grid tapes need nwscr enlarged to fit — that is independent
  of this change.

No new dependencies, no interface changes; the two loop replacements
and two scratch integers are the entire diff.