Replace Hermite-Legendre JFNK with structured split solver

[joglekara]

Closes #343.

Summary

• add a Strang/Cayley split integrator with exact streaming and collision half-steps plus structured local Hermite and Legendre force solves
• add midpoint-field correction, minimum-L2 conservation correction, and boundary/high-mode/residual diagnostics
• update the production bump-on-tail configuration to dt=0.1, Nh=32, Nl=128, and the validated [2,18] Legendre window
• retire the global JFNK implicit solver, its configuration surface, documentation, and expensive integration test

Validation

• 23 Hermite-Legendre tests pass; dense reference tests cover both Cayley implementations
• all pre-commit hooks pass
• NERSC t=700 production run completed 7000 bounded steps in 16.57 solver seconds and 55.13 wall seconds with peak GPU memory 1.195 GB
• relative invariant drift stayed below 1.39e-14; final boundary occupancy was 5.10e-5 and high-Legendre fraction was 5.29e-6
• dt=0.05 and Nl=256 refinements reproduced frequency, growth rate, and saturation within about 1.2 percent
• compared with the former JFNK baseline, the split path is about 35x faster per step, 177x faster per simulated time, and uses about 54x less GPU memory