fix(petab): refuse a U-tagged free parameter on export instead of writing it as a hard-bounded one (#736)

[wshlavacek]

Closes #736. Third and last of the export-drops-part-of-a-declaration family, after #719
(PR #734) and #733 (PR #735).

What was wrong

uniform_var = v 0 10 U (and its loguniform_var twin) means the box is enforced only
during initialization — it seeds the first population and the search is then free to leave
it. ploop reads the tag and the config loader honours it, so the fitter builds a parameter
whose box is (-inf, inf).

The exporter read value[0] and value[1] and stopped. The comment above the construction
called the third token "the native bounded flag, inert for the location families" — true
there, and false where the token is actually legal: parse.b_var_def_keys is exactly
['loguniform_var', 'uniform_var'], the two families whose p1/p2 become the PEtab
bounds. So bounded defaulted to True and _petab_uniform_row wrote [p1, p2] into
lowerBound/upperBound, which in PEtab are hard box constraints.

Measured end to end:

1. fitter:       v1.bounded = False   box = (-inf, inf)
2. exported row: v1	true	0	10              # byte-identical to the bounded spelling
3. re-imported:  v1.bounded = True    box = (0.0, 10.0)
   conf line: ['uniform_var = v1 0 10']         # the U is gone

A search the config declared unconstrained comes back constrained, silently — against
docs/petab.rst's "the export is fit-preserving" and against the module's own contract that
everything it cannot write raises NotImplementedError.

Why a refusal and not a better mapping

Unlike #719 and #733, there is no mapping waiting to be written. PEtab's nearest shape — blank
(infinite) bounds plus an explicit priorDistribution = uniform over the box — says something
else, because PEtab bounds truncate a prior rather than seed a draw. It does not survive the
trip either:

row = PetabParameterRow('v1', True, None, None,
                        prior_distribution='uniform', prior_parameters=(0.0, 10.0))
free_parameter_from_row(row).bounded      # True

_resolve_prior's uniform arm returns bounded=True unconditionally. That is the right
reading for a PEtab row, whose bounds are hard, so nothing changes on the import side.

Refuse vs. warn-and-export. I considered writing the box as bounds with a warning and
rejected it: the exporter emits no warnings anywhere, so it would be a new channel for one
case, and a warning is precisely the signal the #583/#719 work found people do not act on.
The cost of refusing is nothing measurable — there is no U-tagged declaration in the repo's
examples, tests or benchmarks, or in the BNGL-Models/pybnf-jobs corpus. ADR-0025 records the
decision and names warn-and-export as the fallback if it ever becomes a real obstacle.

What changed

• _free_parameter_from_var_line passes the flag through, so the parameter the exporter
  builds matches the one the fitter builds.
• _petab_uniform_row raises NotImplementedError when the box is off, naming the boundary
  and saying to drop the tag to export the box as real bounds.

The untagged and B-tagged spellings export exactly as before.

Tests

Five, of which four fail against the old code.

• Two at the row mapper, parametrized over both flag-bearing families, each also asserting the
  bounded spelling of the same declaration is untouched.
• Two end to end on a real job, one per family.
• One holding the other half of the flag: an explicit B tag exports byte-identically to
  leaving it off — the guard that threading bounded through the constructor perturbs nothing.

Full suite green: 5162 passed, 25 skipped. ruff@0.15.14 clean; the -W --keep-going
Sphinx build succeeds.

Docs

ADR-0025 carries the amendment. docs/petab.rst names the boundary in the export section, and
the uniform_var config-key entry says the tag cannot be exported — where a reader of that tag
will actually look.