Extract the surface-chemistry solver out of m_ibm.fpp into its own module

[sbryngelson]

Follow-up to #1821. src/simulation/m_ibm.fpp is now 2220 lines, and about 400 of them are a surface-chemistry Newton solver that has nothing to do with immersed boundaries.

The seam

The surface-chemistry block (m_ibm.fpp:1822-2218) is:

routine	lines	what it does
s_solve_surface	151	damped Newton with backtracking
s_solve_surface_linear_system	67	dense LU with partial pivoting
s_surface_species_residual	41	per-species flux balance
s_blend_ghost_state	34	bounded ghost reconstruction
s_surface_residual	27	residual assembly
s_surface_energy_residual	18	surface energy balance
s_pick_bath_species	14	sum-constraint closure

Its interface to the immersed boundary is narrow: given (pres, T_IP, T_wall, d, Ys_IP, W_species, thermal_bc) it returns (Ys_s, T_s, mdot_s, converged). Nothing in it knows what a ghost point is. Only the ~70-line branch in s_ibm_correct_state is genuinely IB coupling.

Note s_blend_ghost_state is arguably the exception — it is about the ghost state — but it belongs with the solver whose clamps it exists to preserve, and it takes no IB arguments either.

Performance: this costs nothing

The concern with splitting device code is offload and inlining. It does not apply here:

• m_ibm already makes 11 cross-module device calls from inside the ghost-point GPU_PARALLEL_LOOP (into m_thermochem and m_surface_thermochem).
• 8+ modules already host GPU_ROUTINE(parallelism='[seq]') helpers called from parallel loops (m_collisions, m_igr, m_thinc, m_riemann_state, m_compute_levelset, …).

A new m_surface_chemistry.fpp is the pattern the code already depends on.

The real payoff is testability

Codecov on #1821 reports 56.97% patch coverage, 55 uncovered lines in m_ibm.fpp. The Newton solve is currently reachable only by running a 2-D reacting IB simulation and comparing a golden — which is also why its one regression test is sensitive to the compiler (see #1821, cold-wall golden).

Extracted, it can be tested directly: feed a known surface equilibrium, assert the residual converges to zero; feed a singular Jacobian, assert converged = .false.; assert sum(Ys_s) = 1 holds exactly; assert the ghost blend never returns a negative mass fraction or a temperature outside [T_surface_min, T_surface_max]. None of that needs CFD.

Scope

Pure move plus tests — no behaviour change, so the existing goldens must not shift.

---

Tracked in #1893.

[sbryngelson]

Not started, and worth recording why: this rewrites m_ibm.fpp, which #1821 is still changing, so doing it before that merges guarantees a conflict in the one file both touch.

Independent of the pyrometheus work in #1891 — the two can be done in either order once #1821 is in.

Tracked in #1893.