fix: gradient-fit a piecewise model with a count likelihood

[wshlavacek]

Two independent defects made any gradient fit (lbfgs / trf / gntr) score inf or nan
everywhere on a model built from if()-gated rate laws with a negative-binomial count likelihood
— the shape of every COVID-19 compartmental job in the corpus (Lin 2021, Mallela 2024). Neither is
reachable from a .conf; both are ordinary bugs.

1. Over-requested output-sensitivity selectors

BngsimModel._extract_output_sensitivities asked bngsim for an expression: sensitivity for
every global function whenever print_functions was on. bngsim differentiates function bodies
symbolically and refuses, per function, any body carrying a non-differentiable construct — if(),
a comparison, min/max/abs/floor, a table function (lanl/bngsim#198) — and
Result.output_sensitivities raises ValueError if a refused function is among the requested
selectors.

has_sensitivities_expressions only says "an expression block exists", not "every expression is
differentiable", so it could not stand in for the per-function verdict. A piecewise model has
every function refused — all 14 of the Mallela 2024 model's are if() chains — so every
simulation died with ValueError and the fit's objective was inf at every point.

The request is now filtered to the differentiable expressions
(_differentiable_expression_names, reading bngsim's per-function verdict map). Notes:

• consumers address the tensor by selector name (OutputSensitivities.slice_for), never
  positionally against the Data columns, so dropping a selector is safe;
• stack_scan_sensitivities takes selectors from the first dose point and every point shares them
  (the verdict is a property of the model, not the point);
• a scored column is essentially never one of these functions — the fit observable is an
  observable. If one ever were, its absence now surfaces as a clean "no sensitivity column"
  gradient error instead of a dead simulation;
• no support map (an older bngsim, or a Result loaded from disk) ⇒ nothing to filter on ⇒
  behavior is exactly as before;
• the same filter is applied on the KINSOL steady-state path, which builds the same selector list.
  The SBML path (bngsim_sbml_model.py) requests only species: selectors and is unaffected.

2. Divide-by-zero in the negative-binomial derivatives

d(data_fit)/d(mean) = r (mean − obs) / (mean (r + mean)) is 0/0 → nan (observed zero) or −inf
at a MEAN-centered prediction of exactly 0. That is not a pathological point: any model gated
off over part of the fit window predicts exactly zero there — an epidemic model before its start
time t0, a stimulus-driven readout before the stimulus — and data_fit scores such a point
finitely by clipping prob = r/(r+mean) at 1 − 1e-10. The nan propagated through the
gradient and out to the optimizer, which then tried to assign a free parameter nan:

pybnf.pset.OutOfBoundsException: Free parameter beta cannot be assigned the value nan

_mean_for_slope now mirrors the value path's clip on the mean side (prob ≤ 1 − _PROB_CLIP is
exactly mean ≥ r · _PROB_CLIP/(1 − _PROB_CLIP)), so every expression that divides by the mean
gets the finite slope belonging to the objective actually being scored. MEDIAN centering never
reaches the floor (_mean_for_median is strictly positive), so it is a no-op there.

location_fisher deliberately keeps its mean <= 0 → 0.0 guard rather than adopting the floor,
and now says why: the true I_mean diverges as 1/mean, so flooring it would put a
~1/(r·_PROB_CLIP) weight on every gated-off row and let those rows dominate the EFIM, crushing
the gntr trust-region step. Zero — "this row carries no information about the mean" — is the
stable reading, and the gradient (which does floor) supplies the direction.

Verification

• Full suite: 3338 passed, 9 skipped (rebased on main).
• New unit tests: three for the selector filter (one differentiable + one refused; all refused;
  no support map) and three for the count family at a zero prediction (finite slope and dispersion
  gradient for observed 0 / 1 / 4212; and the slope equals the un-clipped closed form at the very
  mean the value path's clip already pretends to score, so value and gradient stay consistent).
• End to end on the Mallela et al. 2024 NYC job (42 species, 88 reactions, 14 if() functions,
  649 daily counts, neg_bin with a fixed dispersion): before, every start returned inf; after,
  L-BFGS-B converges to nll 5235.80, improving on the published MAP's 5313.95 in the same
  box. The same holds for all four MSAs in that paper — see
  pybnf-jobs/Mallela-2024/*/VALIDATION.md in the BNGL-Models corpus.

Follow-up (not in this PR)

bngsim exposes the per-function verdict only as Result._expression_sens_support; a public
accessor upstream would let this filter drop the private-attribute getattr.