feat(schemaview): TermDescriptor and RangeInfo::is_numeric()

[kervel]

Two additions requested by a downstream consumer that currently reimplements this logic locally
because it is private here. Both are things this crate wants anyway.

1. TermDescriptor — the value→term decision, without a value

The turtle writer spelled out the same precedence chain at three call sites (single value, list
item, mapping value): reference-or-IRI-range → enum meaning → literal. Each copy was ~20 lines of
triple building, and the chain's ordering invariant was stated nowhere.

SlotView::term_descriptor(&Converter) resolves that chain from the slot alone — an enum meaning
map, an IRI disposition, a datatype, a language tag — and returns None when the values are not a
term anything can reproduce (an inlined structure serializes as a blank node whose label nothing
can reproduce). turtle::term_for(&descriptor, value, conv) applies one to a single value.

The three call sites collapse to one line each; turtle.rs is 41 lines shorter.

Writing the chain once also makes an invariant visible that was previously accidental: the writer
tested is_range_iri before the enum meaning, while the order is in fact immaterial, because
determine_rdf_type_info yields (None, false) for an enum range so is_range_iri is never true
for one.

Two deliberate deviations from the shape that was requested:

• term_for is a runtime function, not a TermDescriptor method. schemaview depends on
  neither oxrdf nor serde_json and shouldn't start; the descriptor is schema knowledge, the
  Term is not.
• term_descriptor takes the converter rather than deriving one internally, so enum meaning
  CURIEs expand exactly as they do during serialization. Deriving one would diverge for a schema
  with conflicting prefixes — which this repo has a test for.

2. RangeInfo::is_numeric()

The question a consumer asks when deciding whether a column's values compare as numbers or as
text: '9' >= '10' holds as text and fails as a number, so getting it wrong is silent in both
directions.

is_integer/is_floating_point cannot answer it. They pick a JSON canonicalisation at boxing
time, and between them cover exactly integer, float, double, decimal — so a slot declared
xsd:int, xsd:long, xsd:short, xsd:byte, any xsd:unsigned* or any bounded-integer datatype
reads as non-numeric. Both are left exactly as they are; their callers want the narrower question.

is_numeric uses the same IRI-primary, builtin-name-fallback resolution, so it also catches
schema-defined subtypes and still works against a schema whose linkml:types is not loaded.

Testing

• One case per precedence rule against the descriptor (schemaview/tests/term_descriptor.rs),
  including the enum with no meaning (must stay a literal) and a slot declaring both
  in_language and a datatype (the datatype must win).
• The exact Term per rule through term_for
  (runtime/tests/term_descriptor_parity.rs), on the same fixtures the golden turtle tests use —
  including the mixed enum whose unknown value has no meaning, and a CURIE-valued IRI range.
• The existing turtle serialisation tests pass unchanged, which is what makes the first half a
  refactor rather than a rewrite.
• One slot per XSD numeric datatype for is_numeric, plus the negatives (string, boolean, date,
  uriorcurie, a class range, an enum range) and a schema with no types: block at all, which is
  the case a hand-kept list of datatype IRIs gets wrong.

cargo fmt --all --check, cargo clippy ... -D warnings, cargo test --workspace (73 suites) and
the stub_gen check are all clean.

Not in this PR

• turtle.rs still has two literal_and_type call sites (~lines 555, 611) that make a
  datatype-only version of the same decision for top-level collections of bare scalars emitting
  rdf:value. Routing those through the descriptor would change serialisation output (they would
  start honouring language tags, enum meanings and IRI ranges), so it wants its own change.
• try_lang_tag_collapse — the jsonld:language/jsonld:value class pattern — still collapses an
  inlined object into a language-tagged literal outside the descriptor. The descriptor returns
  None for inlined ranges, so a consumer refuses those columns rather than disagreeing.