KineticForces - FEATURE - Per-bounce-harmonic torque density output

[logan-nc]

Why

The torque density is already computed harmonic by harmonic — the ℓ loop in integrate_psi_quadgk
fills harm_vals and then reduces it — but only the sum was ever kept.

Multi-code NTV benchmarking wants the breakdown. The ITPA TC-24 activity is comparing the
contributions from ℓ = 0, ±1 … ±5 across GPEC, NTVTOK and MARS-K, and the Fortran parent already
emits dTdpsi_fgar(psi, ell), so the Julia code could not answer that request at all.

What

KineticForces/<method>/ gains dTdpsi_ell on a new ell dimension scale. Shape is (npsi, nell)
in Julia and (nell, npsi) to a C-ordered reader, matching the Fortran layout. Summing over ℓ
reproduces dTdpsi. Per-species groups get it with no extra code.

Three things that needed care

• The per-ℓ log is permuted by the same sortperm as psi. QuadGK visits nodes in refinement
  order, not in ψ order, so an unsorted log would misalign with the grid silently rather than
  erroring.
• harm_vals is reused at every ψ node, so the log stores a copy.
• combine_species_states interpolates onto a union grid, and multi-ion runs need per-ℓ on the
  summed total as well as per species. The interpolation is now a shared helper used by both the
  total and each ℓ column, and the combined dtdpsi is taken as the row sum of the per-ℓ matrix —
  which makes the sum rule exact by construction rather than true to interpolation roundoff. A
  species without per-ℓ data leaves the combined matrix empty instead of erroring.

Caveats, documented in the annotation

Quadrature adapts on the summed integrand, so a sharply resonant individual harmonic can be less
well resolved than the total. Profiles still come from the first toroidal mode only, so for
npert > 1 the sum over ℓ matches dtdpsi but neither integrates to total_torque.

Testing

runtests_multiion.jl 64/64 and runtests_kinetic.jl 1845/1845 pass. New assertions cover the sum
rule through combine_species_states, the on-disk shape and orientation, the ℓ-sum on disk, the
dimension-scale attachment, and the mixed case where one species lacks per-ℓ data.

The on-disk orientation was checked from both sides: HDF5.jl reads (npsi, nell) and h5py reads
(nell, npsi), with values indexed as intended.