chartmap: type-bound invert_cart (Cartesian inverse) on fortnum solvers

[krystophny]

What

Add a type-bound cs%invert_cart(x, u_guess, u, geom_status) to the chartmap
coordinate system: the warm single-point Cartesian inverse the full-orbit
pusher needs, ported onto fortnum's released solvers, with the map geometry
owned by the chart and loss policy left to the caller.

• Interior solve: multiroot_hybrid in the pseudo-Cartesian chart
  w = (X, Y, phi) = (rho cos theta, rho sin theta, phi) with the analytic
  Jacobian from covariant_basis (no finite differences). The (X, Y) chart
  is regular through the magnetic axis where the polar Jacobian
  dx/dtheta ~ rho makes det(Jc) -> 0.
• Near-edge solve: a bracketed 1-D radial solve (multiroot_hybrid at
  n = 1) along the seeded (theta, phi) ray, with the bracket [rho_lo, 1]
  enforced in the callback so the radius cannot overshoot past 1. A marker
  leaving the plasma stalls cleanly on the clamped edge instead of grinding.
• Cold multi-seed fallback for seam-stale guesses; the orchestrator keeps
  the lowest-residual root across paths.

Map-specific geometry moves into the chart: chartmap_pseudocart_basis
(polar -> (X,Y) chain rule), chartmap_w_to_u (X/Y axis regularization),
chartmap_radial_scale (|dx/drho|).

geom_status is geometric only (CHARTMAP_LOCATED / CHARTMAP_CLAMPED_EDGE
/ CHARTMAP_OUTSIDE / CHARTMAP_NO_ROOT) and never encodes a confinement loss;
that decision stays with the caller (in SIMPLE, the guiding-centre test).

Why

The full-orbit Boris pusher in SIMPLE hand-rolled a damped Newton with a
backtracking line search to invert the chartmap every microstep. The inverse
is a property of the chart, so it belongs with the chart, and the generic
iteration belongs in fortnum. This PR is the libneo half: a chart-owned
inverse on fortnum solvers. The near-edge bracket removes the line-search
overshoot-into-the-clamp that made near-loss markers expensive. SIMPLE adopts
this in a separate change and keeps its loss policy unchanged.

Verification

fo check: build and tests pass. Full fo pipeline: Static OK (43 modules),
Build OK, Lint OK.

New behavioral test test_chartmap_invert_cart (ctest #48) passes all six
checks: interior roundtrip LOCATED; seam roundtrip with stale phi LOCATED;
near-axis rho recovered via X/Y healing; axis point handled; last-surface point
CLAMPED_EDGE; past-edge target OUTSIDE with rho clamped to 1.

Pre-existing failures, not introduced here: 11 ctest cases
(setup_chartmap_volume, setup_vmec_chartmap_python and the fixture-gated
tests that consume their output) fail because the system Python the test runner
uses lacks the map2disc module (installed only in the project .venv), so the
chartmap/vmec fixtures are never generated. Confirmed identical on pristine
HEAD before any edit; none touch the invert_cart code path. Pointing the
runner at the project .venv is tracked as a follow-up.

Follow-ups (not in this PR)

• SIMPLE side: rewrite invert_cart_warm as a thin loss-policy shell over
  cs%invert_cart (map geom_status -> FO_OK/FO_LOCATE_FAIL via
  accept_or_fail, keep fo_to_gc's guiding-centre loss test, the #427
  contract), then delete the duplicated SIMPLE-side inversion routines. Gate on
  the FO golden-record / W7-X confined-fraction benchmark.
• Optionally surface res_norm/radial_scale from cs%invert_cart so the
  caller classifies without re-evaluating the map.
• Point the libneo test runner's Python at the project .venv so the 11
  fixture-gated chartmap tests run.