[ad-20] add external residual solver example

[krystophny]

Adds a small external residual-solver example over fixed-boundary interior DOFs.

• Builds F(x) from the public raw residual.
• Routes the gradient through State.energy_gradient (exact when the build provides it, raw-force fallback otherwise) and the Newton-Krylov matvec through State.raw_residual_jvp.
• Counts residual, preconditioner, and HVP calls; the FD residual JVP adds its two hidden residual evaluations to the counter, the exact JVP adds none.
• Includes a Krylov root-solver smoke path and a JVP-backed Newton-Krylov step.

Closes #21.

Verification

Test fails before

$ bash -lc 'test -f test-CI/unit-pygvec/test_ad_external_optimizers.py || { echo "missing external optimizer test before AD-20"; exit 2; }'
missing external optimizer test before AD-20

Test passes after

Enzyme-off isolated build (FD path):

$ pytest test-CI/unit-pygvec/test_ad_external_optimizers.py -q --require-pygvec
7 passed in 1.31s

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

$ pytest test-CI/unit-pygvec/test_ad_external_optimizers.py -q --require-pygvec
7 passed in 0.82s

test_exact_hvp_does_not_count_residual_evaluations asserts the counter contract on both paths. Full stack AD suite at this tip: 31 passed on both builds.

[krystophny]

Implemented AD-20 here.

• Added python/examples/external_optimizers.py for fixed-boundary interior residual solves through the public pyGVEC AD API.
• Added tests for raw residual projection, raw-gradient fallback, preconditioner LinearOperator, separately counted FD-HVP calls, Krylov root residual reduction, and HVP-backed Newton-Krylov reduction.

Verification: focused AD-20 pytest passed (6 passed), adjacent AD API/preconditioner/HVP regression pytest passed (26 passed).