Pressure relaxation inverts the stiffened-gas isentrope in closed form; a generic EOS needs a numerical inverse

[sbryngelson]

Split out from #1762, which centralized the isentrope parameters but not the inversion. This is
the one place where a second equation of state needs genuinely new numerics rather than a new
coefficient supplier.

What is there now

s_equilibrate_pressure (src/simulation/m_pressure_relaxation.fpp) drives a Newton iteration on
the equilibrium pressure. Each step needs the density on each phase's isentrope at the current
pressure, and the derivative of that density:

rho_K_s(i) = alpha_rho(i)/max(alpha(i), sgm_eps) &
             *((pres_relax + isentrope_B(i))/(pres_K_init(i) + isentrope_B(i)))**(1._wp/isentrope_n(i))
f_pres  = f_pres  + alpha_rho(i)/rho_K_s(i)
df_pres = df_pres - alpha_rho(i)/(isentrope_n(i)*rho_K_s(i)*(pres_relax + isentrope_B(i)))

Both come from the stiffened-gas isentrope p + B = const * rho**n having a closed-form inverse
rho(p) and a closed-form drho/dp. #1762 made isentrope_n and isentrope_B first-class
(f_isentrope_exponent, f_isentrope_pressure, f_pressure_on_isentrope), so the parameters now
have one definition - but the inversion above is still written for that particular functional form.

Why a generic EOS cannot reuse it

For a Mie-Gruneisen or JWL fluid the isentrope has no analytic inverse. JWL's isentrope

$$p_s(\rho) = A e^{-R_1 \rho_0/\rho} + B e^{-R_2 \rho_0/\rho} + C (\rho/\rho_0)^{\omega+1}$$

cannot be solved for rho in closed form, so rho(p) must come from a numerical inversion, and
drho/dp from differentiating the EOS rather than from a power law. That is new solver code, not a
new coefficient table - and it sits inside a Newton loop in a per-cell device routine, so it is a
nested iteration on the GPU with all the convergence and cost questions that implies.

Suggested scope

• An operator pair f_isentrope_density(pres, pres_0, rho_0, i) and its derivative, with the
  stiffened-gas closed form as the default implementation so nothing changes today.
• A bracketed Newton (or Ridders) fallback for EOS families with no analytic inverse, with an
  iteration cap and a documented failure mode - it must not silently return a non-converged density
  inside the outer Newton solve.
• Convergence and cost measured before it is enabled: this is a per-cell device routine in the 6-eq
  path, and a nested iteration there is the kind of change that regresses the hot loop. Pass an eos_state to s_compute_speed_of_sound #1714 was
  abandoned over a 20% regression from a much smaller change to this area.
• A manufactured test with a known non-analytic isentrope, so the fallback is exercised without
  needing JWL to land first.

Until then, case_validator.py should refuse model_eqns = 3 pressure relaxation for any fluid
whose EOS lacks an analytic isentrope inverse. Blocks the Mie-Gruneisen step of #1638 only for
6-equation cases; the 5-equation Allaire path that #1638 actually targets does not reach this code.