fix(objective): ave_norm_sos's normalizer averages the measured points only (#707)

[wshlavacek]

Closes #707.

The defect

AveNormSumOfSquaresObjective.evaluate precomputed each column's normalizer with
np.average over the raw column, NaNs included. A NaN in an experimental column is
missing data and a first-class supported value — the reader parses it natively,
docs/config.rst tells users to write it, SummationObjective.evaluate skips those rows,
and #479 already made every Data normalization reduce past them, because a sparse
multi-observable column carries NaN wherever that observable was not measured.

But this mean is a divisor. Over the raw column it was NaN, so ((sim - exp) / ybar)**2
returned NaN for every present point of the column too, and the whole objective went NaN.

Nothing downstream recovered it. evaluate gates the prediction for NaN/inf but not the
accumulated func_value, so it returned a float rather than None — a NaN is not a failed
simulation, so it never took the failed-simulation path and reached the optimizer as a real
score. Every NaN comparison is False, so new_score < best_score was never true: every
parameter set was silently rejected, and the fit ran to completion reporting no improvement,
with no error and nothing in the log to say why.

Scope: four sites, not two

The issue named two. Four computed this same column mean, and all four had it:

site	consequence
pybnf/objective.py — legacy objfunc = ave_norm_sos	whole objective NaN
pybnf/noise/source.py — ColumnMeanSigma, i.e. modern objective = ave_norm_sos	same
pybnf/petab/export.py	writes the literal nan as an observable's noiseFormula — not a usable PEtab problem
pybnf/petab/import_.py	compares a constant sigma against the column mean to tell ave_norm_sos from sos; NaN compares equal to nothing, so a round-tripped ave_norm_sos came back silently downgraded to sos

The two PEtab sites are not hypothetical: PEtab's own round-trip writes nan for an
unmeasured point (import_.py:1219), so a sparse column is exactly what arrives there.

ColumnMeanSigma.value's comment claimed its result was identical to the legacy path. It
was — poisoning included. All four now read one pybnf.data.observed_mean, so that claim is
structural rather than two parallel implementations that happened to agree.

Behaviour

On a dense column observed_mean is the plain mean, so every existing fit and golden value
is byte-identical. A column with no measured point has no mean and gives NaN; harmless,
since every one of its rows is skipped in scoring and the normalizer is never read. It is
written out rather than using np.nanmean so that case returns quietly instead of warning
Mean of empty slice once per unscored column on every evaluate() of the fit.

Tests

The NaN-in-experiment case was untested for this objfunc — the one existing ave_norm_sos
NaN test puts the NaN in the simulation. Added:

• test_objective_funcs.py — the sparse-column value for the legacy objfunc, and the
  wholly-unmeasured-column case (which also asserts no warning is raised)
• test_objective_surface.py — the same for the desugared modern surface
• test_petab_export.py — the exported noiseFormula on a sparse problem
• test_data_class.py — two unit tests for observed_mean, beside Feature request: composable floor-normalization + analytic per-series scaling for relative / arbitrary-unit data #479's nan-aware test

All four behavioural tests fail against the old code (verified by reverting the four
source files and re-running).

Full suite with BNGPATH set: 5093 passed, 25 skipped. Docs rebuilt clean under
-W --keep-going.

Docs

algorithms.rst described ybar as "the average of the entire data column", which is now
wrong, and the objective-function preamble did not mention that missing points are skipped at
all. Both now say what the code does, as do the two column_mean noise-source descriptions
in config_keys.rst and noise_models.rst.

Not in this PR

While confirming the above I found the same hazard in three normalizations #479 did not
reach — normalize_to_zero NaNs a whole sparse column, _subtract_baseline does too when
row 0 is unmeasured, and normalize_to_unit_scale then crashes with
ValueError: All-NaN slice encountered. Separate code, separate config keys, and the
_subtract_baseline fix needs a call on what "baseline row" means for a sparse column, so I
left it out rather than widen this. Filing separately.