Drop bundled MINPACK; use fortnum multiroot for quasi symplectic steps

[krystophny]

Merge order

Merge sequence: #390, then #391 (#391 stacks its test coverage on this PR).

Both PRs are based on main individually, so each diff is cumulative against
main and the two overlap. Merge #390 first. Once it lands, the #391 diff
shrinks to its increment: the test_sympl_stell coverage for the fortnum
multiroot quasi path.

Scope

orbit_symplectic_quasi was the only consumer of the vendored MINPACK hybrd1. This PR removes the bundled MINPACK and routes all eight root-finding call sites through fortnum's multiroot_hybrids (Newton step with a backtracking line search and a finite-difference Jacobian), preserving the warm-started, Jacobian-free solve the call sites relied on. Each f_*_quasi residual now conforms to the multiroot_fn_t interface (assumed-shape x/f, optional ctx, no iflag); residual state still travels through the existing module globals f, fs, si, and the f_midpoint_quasi midpoint-then-endpoint double field evaluation is unchanged. xtol and ftol are both set to si%rtol to track the former hybrd1 stopping behavior. fortnum is wired via CMake FetchContent (GIT_TAG a7faa3c, guarded by NOT TARGET fortnum) and linked into the simple library.

Dependency removed: bundled MINPACK (src/contrib/minpack.f90, src/contrib/minpack_interfaces.f90, src/contrib/LICENSE.minpack).

Globalization note

hybrd1 used a Powell hybrid (dogleg trust region); multiroot_hybrids uses a Newton step with an Armijo backtracking line search. Both are warm-started from the previous step and use a finite-difference Jacobian. The direct symplectic-integrator tests below pass within their existing tolerances, but the full golden-record trajectory comparison could not be run in this environment (see Unverified).

Verification

Configure and build (fortnum fetched at a7faa3c, heap-trampoline path available):

$ cmake -S . -B build_fortnum -GNinja -DCMAKE_BUILD_TYPE=Fast -DENABLE_PYTHON_INTERFACE=OFF
-- Performing Test FORTNUM_HAVE_HEAP_TRAMPOLINES - Success
-- Configuring done (16.8s)
-- Generating done (0.5s)

$ cmake --build build_fortnum --target simple simple.x -j32
[562/562] Linking Fortran static library src/libsimple.a
[4/4] Linking Fortran executable simple.x      # exit 0

orbit_symplectic_quasi.f90 compiles with no warnings; the full target set links (427/427).

Direct symplectic-integrator tests (these drive euler1, euler2, midpoint, and gauss4 quasi steps through the new solver with rtol=1e-12 and assert energy-drift tolerances):

1/75 Test  #1: test_sympl_tokamak ......................   Passed    0.08 sec
3/75 Test  #3: test_sympl_testfield ....................   Passed    0.10 sec
6/75 Test  #2: test_sympl_tokamak_plot .................   Passed    2.72 sec
15/75 Test #4: test_axis_crossing ......................   Passed    7.49 sec
24/75 Test #27: test_orbit_symplectic_base .............   Passed    0.11 sec

Full 32-particle Gauss-mode (integmode=3, relerr=1d-13) orbit run completes and writes valid NetCDF output:

$ ./build_fortnum/simple.x        # in examples/orbit_output_test
INFO: Parallel particle tracing completed      69.763 s
 Total field evaluations:                692234
INFO: Output writing completed       0.057 s

$ ./test_orbit_netcdf_output.x    # verify orbits.nc
 PASS: Time data contains valid values
 All NetCDF orbit output tests PASSED!

51 of 75 ctest cases pass. The remaining cases fail for environment reasons unrelated to this change:

• 9 golden_record_* cases: the harness builds a second reference copy of SIMPLE that clones fortplot over HTTPS, which fails here (fatal: fetch-pack: invalid index-pack output, Failed to clone repository: fortplot.git). The failure is in the reference sub-build, before any physics runs.
• orbit_netcdf_output (120s budget), test_chartmap_pipeline (60s budget): timed out only under -j8 load. Run standalone they pass (simple.x 84s, test_chartmap_pipeline.x 82s, both exit 0).
• orbit_netcdf_verify, orbit_netcdf_plot: depend on the truncated orbits.nc from the timed-out producer; they pass against a complete file (shown above).

Unverified

The golden-record trajectory comparison (golden_record.sh) builds a reference SIMPLE from main and clones libneo and fortplot over HTTPS; the network is isolated in this environment, so that comparison did not run. Marked draft until it runs against the reference build.

Note: fortnum pin updated to current main (92de6e9) after a fortnum history rewrite; old shas no longer resolve.

[krystophny]

Superseded by #392 (commit b3fe016 on main), whose squash merge already landed the bundled-MINPACK removal and fortnum multiroot migration. Closing as already-merged; this branch's tip is now behind main on cmake/Util.cmake (pre env-cleanup), so merging it would regress.