perf: rebuild presolve as an incremental delta-based bake

[rasros]

Part of #937.

What changed

• Presolve runs on a persistent PresolveSession: a stable-id factor set with tombstones and one persistent PropagationState, instead of rebuilding a Problem per firing pass.
• Passes emit an explicit delta (drop-by-id, add factors, tighten domains); the session applies it and re-propagates incrementally via the dirty-var/watcher machinery, never a from-scratch computeBaked.
• Bake-time probing moved out of the Problem kernel into presolve.RootBaker behind a seedDeductions seam; the six probe flags become a BakeConfig.
• Problem gains a preFolded pass-view mode that defers propagators/occurrences/baked; the solver Problem is materialized once at the end.
• Affine and dup-columns read the session's incrementally-maintained int occurrence index instead of rebuilding it every round.

Why

Presolve rebuilt a fresh Problem (occurrence invert + full bake) after every firing pass, so cost scaled with passes x factors; ma-path-finding took ~1.6s at budget=0.

Testing

• Byte-identical presolve oracle over the 130 mzn-bench instances (dry-run-presolve, budget=0): factor counts, passes fired, folded domains and proven-infeasible are identical to the from-scratch path.
• Bench-parity solve on the budget-truncated instances (cp-single): same optimum / status as presolve off.

Notes

• 123/130 presolve under 0.5s at budget=0; prop_stress drops 7.5s to 0.2s. Seven large instances stay 0.5-1.5s, bounded by inherent per-pass work (subsume and affine eliminations, wide-domain re-propagation), not the bake.
• The feat: bound the presolve phase with a configurable wall-clock budget #945 presolve budget is left in place as the backstop for those seven; removing it is a follow-up gated on making the passes themselves incremental across rounds.