Extract shared Gauss implicit-RK/Newton core for GC, CPP, and full orbit

[krystophny]

What

Extract the shared implicit Gauss-collocation residual / Jacobian / Newton / linear-solve core now used by the guiding-center (GC) integrator, the Classical Pauli Particle (CPP) pusher, and the curvilinear full-orbit pusher into a new device-callable module orbit_rk_core.

This is the third step of the full-orbit stack and the genuine second/third occurrence: GC (orbit_symplectic) and CPP (orbit_cpp) carried byte-identical degenerate-Lagrangian Gauss residual and Jacobian bodies on field_can_t (the CPP guiding-center motion is the slow manifold of the Pauli particle, so the residuals coincide at fixed mu); CPP and full orbit each carried their own dense LU solver.

How

New orbit_rk_core, all routines !$acc routine seq, fixed-size args, integer model dispatch (MODEL_GC, MODEL_CPP, MODEL_FO_SYMPL), no procedure pointers / class() in the hot path:

• rk_gauss_residual / rk_gauss_jacobian -- select case(model) to one inlinable kernel for the field-canonical Gauss step (arithmetic lifted verbatim from f_rk_gauss / jac_rk_gauss).
• rk_solve(n, A, b, info) -- fixed-size partial-pivot LU, replaces dgesv on the device and the duplicated lu_solve / lu_solve6.
• rk_gauss_newton -- the newton_rk_gauss control flow with rk_solve and a returned hit_maxit flag, so the EVT_RK_GAUSS_MAXIT counter stays a CPU-only concern (mirrors how the GPU newton1 path omits fort.6601).

orbit_symplectic's f_rk_gauss / jac_rk_gauss become thin MODEL_GC wrappers; the GC CPU Newton keeps dgesv so its results do not change. orbit_cpp drops its residual/Jacobian/LU/Newton copies and calls the core with MODEL_CPP. orbit_full's foimpl_step calls rk_solve. coeff_rk_gauss (plain data) stays in orbit_symplectic_base.

Net -40 lines across existing modules; the only addition is the new shared module.

Hard constraints honored

• src/simple_gpu.f90 and src/orbit_symplectic_euler1.f90 untouched: the production GPU path is symplectic-Euler + Boozer (EXPL_IMPL_EULER), not the Gauss path.
• coeff_rk_gauss data and field_can_t layout unchanged.
• GC orbit_timestep_sympl_rk_gauss signature and numerics unchanged.

Verification

Regression gate: GC golden / symplectic tests before and after, plus the GPU bench.

The standard Fast profile uses -ffast-math -ffp-contract=fast, which reassociates floating-point freely and differently across a module boundary, so moving identical arithmetic into a new compilation unit shifts the last bit. Measured drift is round-off only: gauss4 1 ULP (4.4e-16), gauss2 2.3e-13 accumulated over 10000 steps. To prove the extraction is arithmetically exact (same operations, same order), the byte-identity check is run under the repo's deterministic-FP profile (-DSIMPLE_DETERMINISTIC_FP=ON, IEEE strict), where a pure code-motion refactor must be byte-identical.

Deterministic-FP make build-deterministic CONFIG=Fast, test_sympl_stell:

# BEFORE (base feature/full-orbit-cpp-symplectic)
e3699b112bbc84908f68e2aa047fb6fb  gauss2.out
60b6786bc1ff40cda318a8337c2f1c03  gauss4.out
a72edc0dda240a271cca898657c072de  gauss6.out

# AFTER (refactor/shared-symplectic-core)
e3699b112bbc84908f68e2aa047fb6fb  gauss2.out
60b6786bc1ff40cda318a8337c2f1c03  gauss4.out
a72edc0dda240a271cca898657c072de  gauss6.out

$ diff before.md5 after.md5   # (empty) -> BYTE-IDENTICAL

Full suite, standard Fast config, after refactor:

$ ctest --test-dir build -R 'sympl|cpp|fo_symplectic|gpu_orbit|orbit_model|full_orbit'
100% tests passed, 0 tests failed out of 11
  test_sympl_tokamak ........... Passed
  test_sympl_stell ............. Passed
  test_fo_symplectic ........... Passed
  test_orbit_model_dispatch .... Passed
  test_cpp_equals_gc_largestep . Passed
  test_cpp_invariants .......... Passed
  test_gpu_orbit_bench ......... Passed   (loss-step mismatches = 0)

$ ctest --test-dir build -LE regression -E 'test_sympl_stell|golden'
100% tests passed, 0 tests failed out of 65

test_gpu_orbit_bench (the OpenACC EXPL_IMPL_EULER path) reports zero loss-step mismatches before and after, confirming the GPU production path is undisturbed.

[krystophny]

Superseded by #429, which landed the keep-worthy nucleus (the GC Gauss/Newton core in orbit_rk_core) rebased onto current main. The CPP and symplectic-FO parts this PR also carried are retired: CPP per #420, and the full-orbit path is the explicit Cartesian Boris (orbit_model=7, #426), not the shelved symplectic prototype.