perf(ode-grad): hoist state-Jacobian eval out of per-parameter loop (gh#400)

[vsbuffalo]

Cuts the per-step cost of the ODE forward-sensitivity gradient (the NUTS-on-ODE wall for long-burn-in fits) by removing a redundant expression evaluation. First of the two cost-per-step levers in gh#400; the second (kill per-step allocation) is deferred — see below.

What

In sensitivity_derivs, the state-Jacobian term ∂rate_r/∂x_j (eval_deriv_entry) was re-evaluated inside the for p in 0..d loop over estimated parameters, even though its value is independent of the parameter column p — only the S[j,p] multiplier varies. This evaluates each transition's J_x entries once (during the first column) and reuses them across the remaining columns, cutting the J_x expression-eval work from P·nnz to nnz per transition per RK4 stage. The force-of-infection rates (dense J_x, ~8–9 nonzeros) benefit most.

Byte-neutral

The gradient is bit-identical before/after. The mechanism matters: column 0 keeps the exact original J_θ-then-J_x evaluation order, so the shared binding cache is populated identically; columns p>0 previously re-called eval_deriv_entry against the warm cache and got exactly the values now stashed. A fully-hoisted variant that evaluated all J_x before J_θ reorders cache population and perturbs one gradient component by 1 ULP — the column-0 interleave avoids that.

A/B gate (det_grad on the seir_observations FD-oracle fixture, f64::to_bits):

baseline  ll=c0598ed08da0c428
          grad=c05a4bbe00ed4375 4070de40d6693ee5 bffbd780d1241668 405e2ba1ec971515
after     identical — all four grad words + loglik bit-for-bit

Verification

• FD gradient oracles green (3/3)
• make test green (all phases: Rust 907, inference, OCaml, integration)

Deferred: lever 2 (per-step allocation)

The companion optimization — threading a reusable Rk4Workspace through rk4_step to eliminate per-step heap allocation — was implemented and reverted. It is not byte-identical: pure storage relocation drifts one gradient component by 1–2 ULP (value trajectory provably unchanged; only the dense-binding sensitivity chain moves), revealing that the forward-sensitivity accumulation is ULP-fragile to RK4 buffer structure via FP-order sensitivity. Held pending a decision on whether to accept a numerically-equivalent (FD-oracle-validated) change or root-cause the fragility first. Tracked in gh#400.