perf: fold wide-sparse domain reductions as compact survivor sets

[rasros]

What changed

• Record a wide-but-sparse bake reduction as its surviving-value set instead of one interior hole per excluded value, in the bake diff, the domain fold, and the seed apply. All three are O(survivors), not O(span).
• Add restrictIntToSurvivors and propagateAtomsForSetRestriction, which wake the affected order/eq atoms off the value-sorted atom index in O(survivors + eq-atoms).
• Gate the survivor form on a span past bitsetThreshold; narrower reductions keep the plain hole path, whose holes are few and which carries the pre-existing-seed dedup.
• Prefer a bitset over a survivor array when it is memory-neutral, for O(1) contains.
• Move EmptyIntArray / EmptyLongArray from Factor to util.

Why

liner-sf-repositioning finishes presolve over 1s because the bake folds Element-GAC-reduced domains through SurvivorsDomain.forEachHole, which is O(span) over domains spanning millions for a handful of survivors (#938). Recording the survivors directly drops liner-sf presolve to ~0.76s and proteindesign/is/tdtsp by 5-11%.

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 and oocsp still prove their exact optimum with identical assignments; liner-sf's anytime incumbent differs only by search-path divergence.