Replay pr3 state rollback and test

[ElyesAhmed]

Summary

This PR completes, hardens, and expands the state rollback mechanism initiated in [#7198].

When a simulator timestep fails to converge and is chopped into smaller sub-steps, any dynamic state modified during the failed iterations or updated at the start of the timestep must be restored to its exact beginning-of-timestep values. Without proper rollback, retried sub-steps start from polluted state, causing non-deterministic trajectories, trajectory divergence, or convergence degradation.

This PR establishes complete state rollback across Blackoil, Well, Aquifer, and Material Law (Hysteresis) subsystems and adds an extensive deterministic timestep replay regression test suite.

---

Key Additions Beyond #7198

1. Relative Permeability & Capillary Pressure Hysteresis Rollback:

   • In conjunction with the companion opm-common PR, dynamic directional hysteresis inflection points and historical saturation extrema ($S_{w,\text{max}}$, $S_{g,\text{max}}$, etc.) are now captured at beginTimeStep() and restored on updateFailed().
   • Prevents irreversible shift of hysteresis scanning curves during divergent/chopped Newton iterations.

2. Linear Solver & Preconditioner Invalidation Documentation:

   • Verified that ISTLSolver::eraseMatrix() completely deallocates AMG hierarchies, smoothers, and coarse matrix factorizations on updateFailed().
   • Documented the role of --cpr-reuse-setup=0 in replay test harnesses: isolates physical state verification from heuristic preconditioner reuse cadences (solveCount_), ensuring exact bitwise replay identity.

3. Expanded Replay Regression Test Suite & CMake Integration:

   • Added tests/run-all-state-rollback-tests.py providing automated verification across 17 distinct physics configurations.
   • Registered high-impact replay regression tests in regressionTests.cmake (runnable via standard CTest):
     • WCYCLE-1 (Well cycling timers rollback)
     • 2_MSWWELL_MODEL3 (Multisegment well state isolation)
     • MOD4_GRP (Group hierarchy and NUPCOL rate allocations)
     • EXTREME_ROLLBACK_STRESS (Multi-physics stress under forced cuts)
     • FORCED_ROLLBACK_MULTICUT (Cascading multi-cut step retries)

4. Architectural Cleanup of Snapshot Structures:

   • Encapsulated snapshot members into dedicated PrevTimestepState structs across FlowProblem, FlowProblemBlackoil, and BlackoilWellModelGeneric.
   • Cleanly excluded transient step snapshots from restart serialization.

---

Subsystem Audit & Rollback Summary

Subsystem / Feature	Snapshot / Rollback Mechanism	Status
Wells & Groups	BlackoilWellModelGeneric::PrevTimestepState (wgstate, nupcol_wgstate, closed_this_step, well_open_times, well_close_times, genNetwork_)	Verified
Rock Compaction	FlowProblem::PrevTimestepState (rockCompTransMultVal_, minEffectivePressure_, maxWaterPoreVolume_)	Verified
Relperm Hysteresis	EclMaterialLawManager::captureBeginTimeStepState() / restoreBeginTimeStepState()	Verified
Polymer Adsorption	FlowProblem::PrevTimestepState (maxPolymerConcentration_)	Verified
ACTIONX / UDQ	Evaluated strictly at converged step boundaries (endTimeStep()), zero mid-iteration leakage	Verified
Dynamic Aquifers	Cumulative influx and pressure tables commit strictly at endTimeStep()	Verified
Solvent / Foam / Brine	Instantaneous evaluations of primary variables; zero cross-step state leak	Verified

---

Companion PR & Dependencies

• opm-common Dependency: Requires companion opm-common PR [https://github.com/Implement hysteresis state capture and rollback in material law manager opm-common#5327] for EclMaterialLawManager hysteresis snapshot APIs.

---

Verification Results

All 17 active physics regression cases in run-all-state-rollback-tests.py achieve bitwise machine-precision identity ($2 \times 10^{-14}$) between adaptive baseline runs (with failed iterations and chopped steps) and hardcoded replay runs:

• SPE1CASE2_ROCK2DTR (16 cut steps) — MATCH (2e-14)
• 0_BASE_MODEL2 (14 cut steps) — MATCH (2e-14)
• WCYCLE-1 (4 cut steps) — MATCH (2e-14)
• 2_MSWWELL_MODEL3 (12 cut steps) — MATCH (2e-14)
• ACTIONX_GCONPROD (15 cut steps) — MATCH (2e-14)
• ACTIONX_UDQ (21 cut steps) — MATCH (2e-14)
• MOD4_GRP (8 cut steps) — MATCH (2e-14)
• 2D_OW_CTAQUIFER (5 cut steps) — MATCH (2e-14)
• 2D_FETKOVICHAQUIFER (4 cut steps) — MATCH (2e-14)
• 2D_THREEPHASE_POLY_HETER (7 cut steps) — MATCH (2e-14)
• 7_HYSTERESIS_MODEL2 (6 cut steps) — MATCH (2e-14)
• WAGHYSTR-01 (9 cut steps) — MATCH (2e-14)
• VAPPARS-01 (8 cut steps) — MATCH (2e-14)
• EXTREME_ROLLBACK_STRESS (18 cut steps) — MATCH (2e-14)
• FORCED_ROLLBACK_MULTICUT (24 cut steps) — MATCH (2e-14)

---

Out of Scope / Deferred for Future Work

• Compositional Model (flow_comp): State rollback hooks for compositional fluid systems.
• GPU ISTL Solver Backend: Reset parity checks in ISTLSolverGPUISTL.