Drop GSL dependency from collision_freqs

[krystophny]

Closes #285.

collision_freqs linked GPLv3 GSL with PUBLIC visibility for one routine, the
lower incomplete gamma in the Chandrasekhar function. The GPL obligation reached
every MIT-licensed downstream consumer (SIMPLE, NEO-2, NEO-RT, MEPHIT, KAMEL,
rabe), including SIMPLE's published binary wheels.

Changes:

• gsl_sf_gamma replaced by the Fortran 2008 intrinsic gamma().
• gsl_sf_gamma_inc_P replaced by an in-tree regularized lower incomplete
  gamma, written from scratch: DLMF 8.11.4 series for x < a + 1, DLMF 8.9.2
  continued fraction (modified Lentz) otherwise.
• find_package(GSL REQUIRED) and GSL::gsl removed from
  src/collisions/CMakeLists.txt; libgsl-dev removed from CI and
  setup/debian.sh; gsl removed from flake.nix.
• GSL stays as an optional test oracle: with find_package(GSL QUIET) found,
  test_incomplete_gamma_gsl_oracle.x compares the in-tree function against
  gsl_sf_gamma_inc_P * gsl_sf_gamma at 1944 points (a in [0.25, 6], x in
  [1e-8, 100]) with relative tolerance 1e-12. Without GSL the test is skipped
  and nothing links it.
• test_collision_freqs.x gains a reference-value test for
  lower_incomplete_gamma against 32 mpmath values (40-digit precision),
  relative tolerance 1e-12, so the function is verified on systems without GSL.

The existing calc_perp_coll_freq expected values, produced by the GSL-backed
implementation, pass unchanged.

Verification

Before (main): configure without GSL fails.

$ cmake -S . -B build_nogsl -G Ninja -DCMAKE_DISABLE_FIND_PACKAGE_GSL=TRUE
CMake Error at src/collisions/CMakeLists.txt:4 (find_package):
  find_package for module GSL called with REQUIRED, but
  CMAKE_DISABLE_FIND_PACKAGE_GSL is enabled.  A REQUIRED package cannot be
  disabled.

After (this branch): same configuration succeeds, build and tests pass without
GSL.

$ cmake -S . -B build_nogsl -G Ninja -DCMAKE_DISABLE_FIND_PACKAGE_GSL=TRUE
configure exit: 0
$ ctest -R "collision|incomplete"
1/1 Test #19: test_collision_freqs .............   Passed    0.30 sec
100% tests passed, 0 tests failed out of 1

After, with GSL present (Homebrew GSL 2.8): oracle agrees with the in-tree
implementation.

$ ctest -R "collision|incomplete_gamma|transport" --output-on-failure
1/3 Test #19: test_collision_freqs ...............   Passed    0.28 sec
2/3 Test #20: test_incomplete_gamma_gsl_oracle ...   Passed    0.26 sec
3/3 Test #21: test_transport .....................   Passed    0.24 sec
100% tests passed, 0 tests failed out of 3

GSL appears on no link line except the optional oracle test:

$ rg -n "gsl" build.ninja | grep -v incomplete_gamma_gsl_oracle | grep LINK
(no output)

Full suite: 56/70 pass; the 14 failures (chartmap/map2disc, tilted_coil) fail
identically on unmodified origin/main in this environment (missing optional
map2disc Python module) and are unrelated to this change.

[krystophny]

@marjohma this is high prio, as it affects licensing of all our codes. As soon as they link to GSL, they must all be distributed under GPL (it is not a catastrophe if we distribute them under MIT, so not an immediate legal problem, but the GPL conditions automatically convert this to GPL per definition if you link GPL software to it, and I prefer to keep MIT).

[krystophny]

Superseded by #316, which routes the collision_freqs lower incomplete gamma to the shared fortnum core (gamma_lower) instead of a libneo-internal reimplementation. Same public interface, same GSL drop; consumers depend on fortnum directly. Leaving this open for a human to close.

[krystophny]

Superseded by #316, which routes collision_freqs to fortnum gamma_lower. #316 builds 362/362 with test_collision_freqs and test_transport green and identical pre-existing failures to main. Closing in favor of #316.