L-BFGS quadratic model through HiGHS

The two-loop recursion is the unconstrained L-BFGS step. HiGHS is used only when that step must sit in a feasible set: an L_inftrust region, a box on x + p, and optional linear equalities.

The QP is min 1/2 ||p - d||^2 with d = -H g and Q = I. A dense compact Hessian is not formed; that matrix is indefinite after a tiny accepted step and Highs_run does not return.

Unconstrained, highs_step is the two-loop direction. A box or trust region scales the whole increment (direction preserved). Packed centering is a mean subtract. Arbitrary equalities still go to HiGHS.

When a host Hessian is present (rgmin_solver_step_hess) and rgmin_solver_set_highs is on, HiGHS solves the convex Newton QP min 1/2 p^T P p + g^T p with per-coordinate boxes from atom maxmove. That is the path eOn / rgpot / any eindir objective share. Without the highs cargo feature the setter returns 1.

Enable the feature (HiGHS is compiled by highs-sys; do that on the remote builder):

cargo test --features highs

use rgmin::{HighsStep, Lbfgs};

let mut opt = Lbfgs::default();
opt.highs = Some(HighsStep {
    trust: Some(0.5),
    lo: Some(-1.0),
    hi: Some(1.0),
    equalities: Vec::new(),
    center_axes: None,
});

landfold --highs is this step on chi, not sequential LP.