fix for composite scaling example

[pascoec]

Closes #1358

1. The symptom. The composite-scaling polynomial example aborted at its own defaults with "EstimateLogP() failure: expected sizeP [9], but got sizeP [7]". It failed on main as well, so it is not a regression.
2. The cause. The early estimate of P splits the digits over whole scaling primes. The real setup splits them over composite limbs, and a digit boundary inside a prime makes the estimate larger than the real value. The check added in Fix BGV ring dimension underestimation with NOISE_FLOODING_MULTIPARTY + HYBRID on 128-bit builds #1158 required exact equality.
3. How big. A sweep of 13,200 composite configurations gave 1,880 mismatches. In 1,872 the estimate was larger, which errs toward security, and in 8 it was smaller.
4. First fix. Making the estimate model the limbs broke 36 working configurations and raised the too-small cases to 51, so it was rejected. Making the check one-sided, throwing only when the real P exceeds the estimate, left every working configuration unchanged and fixed 1,856 of the 1,880.
5. A second bug. The same sweep found 2,505 configurations where context generation never returned. Debugging traced them to three causes in CompositePrimeModuliGen:
   (1) The target prime fell below the cyclotomic order and the unsigned candidate wrapped to just under 2^64.
   (2) Searches started far above the register word and walked down one candidate at a time.
   (3) The raise-the-product loop crossed the register word, turned around and walked to the bottom of the range.
6. Second fix. Throw before the wrap. Start searches at the register word, which picks the same primes. Throw as soon as the raise loop crosses the word. Make PreviousPrime refuse to wrap.
7. A bug the hangs had masked. 222 configurations that used to hang now crashed dividing by zero, because a level held the same prime twice. The adjustment loops removed each rejected candidate from the record of primes in use, so the search could later accept a prime another slot held. Removing only the slot's old prime, once, fixed it.
8. Review cleanup. The search ceiling became 2^registerWordSize + 1, one step lower, which removes one wasted test per clamped search. It changes no selected prime.

Validation.

• Sweep, no hangs: all 13,200 configurations finish, none takes over 60 seconds, and 4,159 now work.
• Sweep, identical primes: every configuration that worked before produces identical moduli.
• Examples and tests: all examples and extras pass, as do the full core, pke and binfhe suites, under both clang 18 and gcc 14.