perf: resume the affine candidate scan from the fold's lowest rewritten factor

[rasros]

Advances #937 (steelmillslab).

What changed

• The affine-elimination candidate scan resumes from the lowest factor id a fold rewrote, instead of restarting at 0 on every elimination.

Why

findAffineCandidate walks the live factors in stable-id order and is called once per elimination, so the loop was O(eliminations · factors). A fold only changes the content of the factors it rewrites (the pivot's occurrences), so no factor below their minimum id can newly become a candidate — everything below is a confirmed non-candidate, unchanged since it was last examined. Resuming the scan from that minimum keeps the exact lowest-id selection order (so the eliminations are unchanged) while turning the loop into O(factors + Σ rewrites). steelmillslab presolve drops from ~490ms to ~410ms (under 0.5s at budget=0); the alias rename rewrites the whole set, so it still rescans from 0.

Testing

• Byte-identical presolve oracle (dry-run-presolve, budget=0) across all 130 mzn-bench instances.
• Bench-parity solve on the budget-truncated instances: ma-path-finding, oocsp, steelmillslab, bus_scheduling all produce an identical solution stream; no wrong-optimum.