KineticForces: preserve periodic vpar spline and enumerate bounce roots exactly

[krystophny]

Summary

• fit the closed-poloidal-turn parallel-velocity factor with PeriodicBC()
• enumerate roots span-by-span, splitting cubic intervals at stationary points
• canonicalize the physically identical theta=0/1 seam
• retain descending root order and existing deepest-well/transit behavior

This ports the relevant endpoint-condition correction from PrincetonUniversity/GPEC#284 without reverting the independent fixes already present from #310.

Why

Wrapping an endpoint-fit cubic closes its value but not its derivatives, leaving a nonphysical seam. The former global root heuristic can also miss a genuine close root pair. The new analytic periodic oracle fails on the old implementation and passes here.

Validation

• kinetic test suite: 292/292 passed on Julia 1.12.6
• analytic periodic seam/root regression: old implementation fails; this branch passes with C2 closure and exact close-pair detection
• TC24 phiI010 L5 full runs: finite TGAR/PGAR results; real-torque change below 3.1e-9 relative
• DIII-D FGAR/TGAR non-regression: unchanged integration counts; real-torque change below 1.1e-6 relative
• corrected Fortran KineticForces: single METHOD_REGISTRY for NTV method flags, docs, and dispatch #284 TC24 archive: completes normally where the old endpoint fit reached LSODE ISTATE=-5
• solovev_kinetic_ntv: all six comparisons pass; the previously missed close bounce pair is found

One unrelated solovev_kinetic_nuzero harness sensitivity remains documented: it occurs before KineticForces is called and varies with adaptive multithreaded ForceFreeStates execution. CI remains the Julia 1.11/formatter authority.