Restore Fourier-space KIM kernels with behavioral tests

[krystophny]

First split item of #175: restore the continuous-Fourier kernels
kernel_rho_phi_of_kr_krp_rg and kernel_rho_B_of_kr_krp_rg as kernels_m
(KIM/src/kernels/kernel_mod.f90), rewired from the removed
plasma_parameter/back_quants globals to the current species_m plasma_t
state, and reactivate the kernel tests as ctest-registered behavioral checks.

Stack: parent branch main, pinned parent commit 2341697e.

Invariant preserved

• Both kernel expressions are restored term-for-term from
  kernel_mod.f90 at 1ae0aeeb~1: Debye branch, large-argument Bessel
  asymptotics, exp(i (k_r - k_r') r_g) phase, 1/(8 pi^2) (rho-phi) and
  i/(8 pi^2 c) (rho-B) normalizations, CGS-Gaussian units.
• The background quantities the kernels consume are computed by species_m
  with the same defining formulas the deleted calculate_backs used
  (A1, A2, lambda_D, z0, ks, kp verified line-by-line);
  Lagrange interpolation keeps the historical binsrc/plag_coeff
  4-point scheme on plasma%r_grid.
• Provider changes only: gsl_sf_bessel_In -> fortnum_special::bessel_in
  (same modified Bessel I_n), and the retired kernel_debye_case config key
  becomes a module switch with unchanged default .false..
• Nothing else links against the new module yet; the existing solver paths,
  golden records, and all 25 pre-existing ctests are untouched.

Test changes

The orphaned KIM/src/tests/ programs pinned values of the deleted
NRL-rounded background (vTe = 4.19e7 sqrt(Te), constant btor); the
current pipeline computes sqrt(T ev/m) and omega_c from B0(r), so
those pins differ at the 1e-3 level while the tightest tolerance was 1e-5:
they cannot pass against either code state. They are replaced by analytic
and property checks at 1e-12 relative tolerance:

• adiabatic Debye screening -1/(8 pi^2 lambda_D^2) at k_r = k_r'
  (the acceptance value named in KIM: revive Fourier-space solver via enforced radial periodicity (localized solutions) #175),
• gyroaverage Gaussian with the Fourier phase at k_r /= k_r',
• k_r <-> k_r' swap conjugacy and r_g phase translation
  (the properties the periodic matrix assembly of KIM: revive Fourier-space solver via enforced radial periodicity (localized solutions) #175 relies on),
• Debye switch removes the force response while reproducing the
  force-free kernel (the historical branch's flipped sign and dropped
  phase are deliberately not restored; see the follow-up comment),
• exact zero of the rho-B kernel without thermodynamic forces and
  linearity of both kernels in A1 and A2.

The dead KIM/src/tests/ directory (empty CMakeLists, broken
kim_init_for_test) and the commented-out KIM_kernels source group are
removed; the new tests live in KIM/tests/ with the other ctest programs.

Verification

Test fails on defective kernel (normalization mutated)

$ sed -i 's|/(2.0d0\*\*3\*pi\*\*2)|/(2.0d0**3*pi)|' KIM/src/kernels/kernel_mod.f90
$ cmake --build build --target test_kernel_rho_phi && ctest --test-dir build -R test_kernel_rho_phi
1/1 Test #23: test_kernel_rho_phi ..............***Failed    0.07 sec
 FAIL adiabatic Debye screening -1/(8 pi^2 lambda_D^2)
   got  =               (-2879.8814430158527,0.0000000000000000)
   want =               (-916.69473434918780,0.0000000000000000)
   rel  =    2.1415926535897900
ERROR STOP

(On main the module and tests do not build at all; the historical test
programs reference modules deleted in 7c1345dc.)

Test passes after restore

$ ctest --test-dir build -R "test_kernel_rho"
1/2 Test #23: test_kernel_rho_phi ..............   Passed    0.06 sec
2/2 Test #24: test_kernel_rho_B ................   Passed    0.06 sec
100% tests passed, 0 tests failed out of 2

Full suite

$ ctest --test-dir build
100% tests passed, 0 tests failed out of 27

Relation to Markus's Fokker-Planck periodic branch

This PR restores the historical continuous Krook kernels. It is useful for Fourier-phase, Debye-limit, and interpolation comparisons, but it is not the off-diagonal Fokker-Planck kernel implemented on feature/kim-forced-periodicity. Hold integration until #191 decides whether any generic tests are retained. The independent FP derivation and oracle are #187, #194, and #196.

[krystophny]

Follow-up commit: the Debye switch now only zeroes the thermodynamic-force terms and keeps the sign, gyroaverage Gaussian, and Fourier phase of the full expression. The historical branch flipped the sign relative to the adiabatic limit and dropped the phase factor, which would anti-screen and break diagonality once the switch feeds a matrix assembly; the modern hat-basis path (artificial_debye_case with the G0 prefactor) uses the retained convention. The switch test now checks that enabling it removes the force response while reproducing the force-free kernel.