gntr assembles forward sensitivities twice per evaluation — DRY the gradient + Fisher-Hessian assemblers into one pass

[wshlavacek]

Follow-up to #481 / #487 (surfaced by a /code-review high pass).

Problem

On the gntr (Fisher/Gauss-Newton trust-region) path, GradientOptimizer.gradient_at calls both:

• assemble_gaussian_gradient — the scalar gradient (_accumulate_experiment walks the scored points), and
• assemble_fisher_hessian — the EFIM Hessian (_accumulate_experiment_fisher walks the same points again).

_accumulate_experiment_fisher duplicates the entire point-iteration scaffold of _accumulate_experiment:

• independent-variable resolution (min(exp_data.cols, ...)),
• the comparable / compare_cols intersection,
• the _sim_row_for row-matching loop + NaN skip,
• the _raw_sensitivity_accessor, and
• the per-point prediction_sensitivity call for dpred_dtheta.

So every scored point's sensitivity row is assembled twice, and _require_gradient_supported is re-run per point on both passes.

Suggested fix

Factor a shared point-iterator (e.g. _iter_scored_points(objective, sim_data, exp_data, routing, ...)) that yields (sim_row, rownum, col_name, weight, dpred_dtheta) once, and have both the gradient and Hessian assemblers consume it — so gntr builds g and H in a single pass and the scaffold lives in one place.

Priority

Low. Correctness-neutral (a pure cleanup/efficiency refactor) and dominated by the bngsim ODE solve, not the PyBNF-side assembly. trf / lbfgs are unaffected (they never assemble H). Filing so it isn't lost.