Qualify hierarchical nonlinear ODE trajectories and higher-order differentiation

[jc-macdonald]

Add the public generic hierarchical workload and correctness evidence requested in #204. The nonlinear two-pool fixture combines shared log rates, group-local rate offsets and local initial conditions, with independent/coupled groups and nonstiff/stiff regimes. Deterministic observations retain a nonzero residual and a documented hierarchical likelihood/prior objective.

Independent scalar cell equations and tight SciPy DOP853 trajectories qualify NumPy/JAX values, conservation, built-in accepted-mesh scans and external scans using only the public fixed-step API. Required small cases check forward/reverse derivatives, Hessians, the gradient of a regularized Hessian log determinant (with positive definiteness verified), reference finite differences, and step refinement. Source and installed-wheel suites execute the same fixture.

References #204. This first PR intentionally does not close the issue: its separate opt-in scaling/fair-comparison measurement PR remains required.

Validation: just format, just ci (839 core tests passed, 389 provider tests passed), just docs, just qualify (NumPy 44, JAX 52, Torch 65; zero selected skips), and just build-all passed. NumPy/JAX required jobs now reject skipped hierarchical cases. No timing thresholds or speedup claims are added. The synthetic model and fixed diagonal regularization are explicitly numerical evidence, not a calibrated application or an unregularized posterior-convexity claim.