Extract shared GC Gauss implicit-RK core (orbit_rk_core); drop CPP and symplectic FO

[krystophny]

Summary

Rebase-and-strip of the shelved shared-symplectic-core work (PR #407) onto
current main. Keeps only the keep-worthy nucleus: the guiding-centre (GC)
Gauss implicit-RK / Newton core, extracted into a new device-callable module
orbit_rk_core. Everything CPP and symplectic-full-orbit is dropped.

What lands

• New src/orbit_rk_core.f90: the GC Gauss-collocation residual, analytic
  Jacobian, a device-portable dense LU solve (rk_solve), and a device-callable
  Newton shell (rk_gauss_newton). Reduced to a single GC kernel; the former
  MODEL_CPP / MODEL_FO_SYMPL codes and their dispatch shells are gone.
• src/orbit_symplectic.f90: f_rk_gauss / jac_rk_gauss become thin wrappers
  over the shared core (MODEL_GC). The arithmetic is moved byte-identically,
  so GC numerics do not change.
• src/CMakeLists.txt: compile orbit_rk_core.f90 before orbit_symplectic.f90.

What is dropped (not merged)

• The CPP pusher (orbit_cpp.f90, MODEL_CPP) and its tests
  (test_cpp_equals_gc_largestep, test_cpp_invariants).
• The shelved symplectic / mocked-Boris full-orbit prototype (orbit_full.f90,
  MODEL_FO_SYMPL), its provider seam (orbit_full_provider.f90) and the two
  mocks, and their tests (test_fo_symplectic, test_full_orbit,
  test_orbit_model_dispatch).
• The branch's params.f90 / simple_main.f90 orbit_model hunks: main
  already owns orbit_model (ORBIT_GC=0, ORBIT_FULL_ORBIT=7) and the
  validated explicit Boris full-orbit pusher (PR Add full-orbit (FO) Boris pusher (orbit_model=7) #426), which this PR leaves
  untouched.

Why

orbit_rk_core is the single home for the GC Gauss step shared by the CPU
integrator and (later) the OpenACC GPU kernel; it is written !$acc routine seq-ready with its own device-portable LU. CPP is retired; the full-orbit path
is the explicit Cartesian Boris (orbit_model=7), not the shelved symplectic
prototype.

Verification

make CONFIG=Fast: build OK. make test: 100% tests passed, 0 failed out of
77 (total 677.80 s). All 14 golden-record regression cases (albert,
albert_coils, boozer, boozer_ncsx, canonical, classifier, classifier_combined,
classifier_fast, collision, meiss, meiss_coils, and the tokamak variants)
pass, confirming the GC numerics are byte-identical after the extraction.

Before (on main): the same 14 golden-record cases pass with the inlined GC
residual/Jacobian. After (this branch): identical results with the arithmetic
moved into orbit_rk_core -- a pure refactor, no numeric change.

Follow-ups (separate PRs)

• Wire rk_gauss_newton / rk_solve onto the fortnum multiroot core in place of
  dgesv for the CPU GC Newton; must preserve the verified GC numerics.
• The device (OpenACC) GC solver path uses fortnum's reverse-communication
  multiroot_step (fortnum PR for the device-capable solver), keeping
  orbit_rk_core as the shared CPU+GPU home.