Drop GSL dependency

[krystophny]

Removes the GSL/FGSL dependency. Replacements come from the libneo math kit
(itpplasma/libneo#290) plus a short in-tree Levin-u summation. All replacement
code is clean-room, implemented from published algorithms.

Resolves #24.

Changes

• src/mephit_pert.f90: the bind(C) interface to gsl_sf_bessel_In_array
  is replaced by neo_bessel_i::bessel_in_array. For pol_mode = 0 the
  I_{m-1} value now comes from the symmetry I_{-n} = I_n instead of an
  out-of-range GSL call.
• src/math_c_api.f90, src/mephit_math.h (new): bind(C) shims
  mephit_integrate_gk (libneo Gauss-Kronrod 21, QUADPACK QAG semantics) and
  mephit_gauss_legendre (libneo Golub-Welsch nodes/weights). No module
  state; the C callback and its context pointer pass through a wrapper local
  to each call, so the shims stay thread-safe.
• src/hyper1F1.c: gsl_integration_qag calls go through
  mephit_integrate_gk; gsl_sum_levin_u is replaced by an in-tree
  levin_u_sum, a 45-line Levin u-transformation after Weniger, Comput.
  Phys. Rep. 10 (1989) 189, eqs. (7.2-8) and (7.3-5).
• src/mephit_fem.c: gsl_integration_glfixed is replaced by
  mephit_gauss_legendre.
• src/mephit_util.c, src/mephit_util.h, src/mephit_run.c:
  gsl_errno_msg and the GSL error handler installation are removed;
  quadrature errors surface as QUADPACK return codes.
• CMakeLists.txt: find_package(GSL REQUIRED) and GSL::gsl removed,
  LIBNEO::math linked, src/math_c_api.f90 added. FFTW is now located
  directly; the libneo math-kit branch no longer leaks FFTW variables into
  the parent scope.
• cmake/SetupCODE.cmake: fgsl-1.6.0 pickup removed.

Depends on itpplasma/libneo#290 for the LIBNEO::math target; draft until
that merges into libneo main.

Verification

Built on macOS (Homebrew gfortran 15, Accelerate LAPACK) against the libneo
math-kit branch.

Before, GSL was a hard requirement and libmephit linked it:

$ git grep -n "GSL" main -- CMakeLists.txt
main:CMakeLists.txt:62:find_package(GSL REQUIRED)
main:CMakeLists.txt:178:  GSL::gsl

After, the full build succeeds without GSL and no binary links it:

$ cmake --build build -j8 2>&1 | tail -4
[183/186] Linking CXX shared library lib/libmephit.dylib
[184/186] Linking C executable bin/mephit_run.x
[185/186] Linking Fortran executable bin/mephit_test.x
[186/186] Linking Fortran executable bin/mephit_post.x
$ otool -L build/lib/libmephit.dylib | grep -i gsl; echo "exit: $?"
exit: 1

The repo has no ctest suite covering the touched paths, so a standalone
harness compares old GSL-backed code against the new implementations. The
old hyper1F1.c (from main, linked against Homebrew GSL 2.8) is compiled
with prefixed symbols next to the new libmephit; GSL is the oracle for
Gauss-Legendre nodes/weights and the Bessel array, matching the libneo
oracle-test pattern (GSL linked only in the test harness, never in shipped
targets). Representative arguments: 25 complex (b, z) pairs for the
hypergeometric routines, quadrature orders 1 to 12, Bessel orders 1 to 16
at six arguments between 1e-3 and 20.

$ ./check_c | tail -2
max rel diff: quad=1.119e-14 accel=1.278e-15 glfixed=4.371e-15
PASS
$ ./check_bessel
bessel In array max rel diff =  8.350E-16
PASS

All differences are at least two orders below the 1e-12 target. The cases
where hypergeometric1f1_quad diverges (integrand singular for the given
b) produce identical signed infinities before and after.

Not verified here: a full MEPHIT run (needs FreeFem++ and run data, not
available on this machine) and the Python tests under test/ (need a
completed run).

[krystophny]

Superseded by #28, which migrates MEPHIT onto fortnum (the single numerical core) instead of the libneo math-kit. #28 has a verified-green full build and numerical-equivalence checks for the Bessel, 1F1 quadrature, Levin acceleration, and Gauss-Legendre paths. Leaving this open for a human to close.

[krystophny]

Superseded by #28, which drops GSL via fortnum directly (Fortran primals + iso_c ABI) instead of libneo math-kit. #28 builds green (91/91), no gsl symbols in libmephit.so. Closing in favor of #28.