Add BakedSatfuncTables: exact baked-table saturation-function evaluation

[hnil]

Bakes each cell's scaled material law (endpoint scaling included, ECL-default three-phase) into minimal exact piecewise-linear tables on three saturation axes, deduplicated across cells. Vertical endpoint scales (maxKr*, maxPc*, e.g. from SWATINIT) are linear in the ordinate and factored into per-cell multipliers instead of keying tables. Node placement is found by midpoint refinement with kink snapping against the installed law, so no endpoint-scaling semantics are reimplemented and the build validates itself (max rel err ~1e-10).

Norne, 44 927 cells, eval3, identical states and checksums:

relperm + pc	ns/cell
generic law stack	89.9
baked	20.6

7 unique tables, 33.7 KB pool for the whole field; 60 B per cell; bake < 0.01 s. Wired into flow it cut Props/update 10.9 → 8.5 s (−23 %) with unchanged iteration counts. Consumer-side hook: OPM/opm-simulators#7304.

Scope of this branch: ECL-default approach, no hysteresis, no directional relperm. A prototype adding Carlson KR hysteresis (second pool for the two hysteretic columns, four scalars per cell, one compare and one add) measures 201.6 → 24.6 ns/cell on Norne with EHYSTR active; it is not in this PR yet. Stone 1/2 should be cheaper still — both put the oil constituents on the sw and sg axes — but is not implemented.

Draft: feedback welcome on where this belongs and how the fallback should be gated.

🤖 Generated with Claude Code

[hnil]

Two follow-up results that bear on how this should be judged.

Hysteresis is where the table representation gains most, not where it struggles. Norne runs EHYSTR ... KR (Carlson). With that active the generic path costs 201.6 ns/cell versus 110.8 without — a 2x overhead that returns identical values at drainage state, because the per-cell parameter object doubles (drainage + imbibition) rather than because any extra mathematics is done. Carlson reduces to imbibition_curve(arg + delta) guarded by one comparison, and only two of the six curves are hysteretic, so a prototype adds a second pool for those two columns plus four scalars per cell:

Norne, eval3, EHYSTR active	ns/cell
generic + hysteresis law	201.6
baked + Carlson, cells at drainage	24.6 (checksum identical)
baked + Carlson, all cells forced reversed	27.7

Pool 33.7 → 49.3 KB, per-cell 60 → 92 B. Not in this branch yet; the imbibition branch still needs end-to-end validation against a reversed state.

A cheaper, independent step exists. Deriving the two-phase arguments once, with no representation change at all, gives 89.9 → 59.8 ns/cell — about a third of the win for none of the machinery here. That is now #5287, and it is worth landing on its own merits regardless of what happens to this PR.

For calibration, the ladder on Norne (eval3, identical checksums): generic 89.9, fused arguments 59.8, baked 20.6; with hysteresis 201.6 → 24.6.