Add reduced Hessian operator and Newton driver

[krystophny]

Adds the exact reduced boundary Hessian and a second-order optimizer driver.

• BoundarySensitivityProblem.reduced_hessian_vector_product: Schur complement H_BB - H_BI H_II^-1 H_IB assembled from exact Enzyme HVPs (two full-space residual JVPs plus one preconditioned interior GMRES solve per matvec). Exact reduced Hessian of the equilibrium-constrained energy at a force-balance equilibrium; the docstring states the validity limit away from one.
• GvecEnergyProblem.hessian_vector_product(x, v): solves the equilibrium at x and applies the reduced Hessian there, with the same failed-solve penalty contract as value and gradient.
• optimize_energy_newton: bounded trust-constr Newton using analytic value, adjoint gradient, and reduced-Hessian hessp; no finite differences anywhere in the chain on Enzyme builds.

Closes no new issue; extends #23/#24 machinery to second order.

Verification

Before

No reduced Hessian existed; second-order optimizers had to build hessp by finite differences of the adjoint gradient (one extra equilibrium solve per Hessian column or matvec).

After

Enzyme-on isolated build (fixed RZ, LLVM/Flang 22.1.6):

$ pytest test-CI/unit-pygvec/test_ad_simsopt.py -q --require-pygvec
11 passed in 2.90s

test_reduced_hessian_symmetric_and_matches_gradient_fd checks <u, H v> = <v, H u> at rtol=1e-6 and the reduced HVP against a centered difference of the adjoint boundary gradient with full equilibrium re-solves (cosine > 0.99, rtol=5e-2, limited by the interior solve floor). test_newton_optimizer_reduces_energy runs the bounded trust-region Newton end to end.

Enzyme-off isolated build (FD fallback path):

$ pytest test-CI/unit-pygvec/test_ad_simsopt.py test-CI/unit-pygvec/test_ad_boundary.py -q --require-pygvec
13 passed, 4 skipped in 30.96s