fix: let Turtle preprocessing own namespace discovery

[jdsika]

Stacked on #57; merge that first and this diff reduces to its own commit.

Summary

• Let the Turtle serializer decide which namespaces earn a declaration, per
  term position, instead of forcing a QName split on every IRI in the graph.
  Splitting unconditionally invented http://a.example/% as the namespace of
  the valid IRI <http://a.example/%25>, and that declaration is not readable,
  so deterministic_turtle refused valid graphs whose subject, object or
  datatype IRI has no prefixed name.
• Traverse triples in lexical predicate order during preprocessing, so the
  generated ns1, ns2, … names stay a function of the graph rather than of
  its insertion order, and present every triple exactly once so the reference
  counts that decide blank-node and collection layout are unchanged.
• Keep the store-order preparation for the formats whose serializers discover
  namespaces while traversing: RDF/XML output is byte-identical.
• Add 61 term-position, binding, keyword-predicate, traversal-count,
  insertion-order and cross-process cases, and document the presentation change.

Presentation change

Turtle-family output no longer declares a namespace that only a subject, object
or datatype used; those terms are written as complete IRIs, so "42"^^ns2:integer
becomes "42"^^<http://www.w3.org/2001/XMLSchema#integer> unless the caller
bound the namespace. One-time diff, documented in the changelog with the
remedy: bind the namespaces you want compact.

Boundary this does not cross

A predicate whose split is not a valid IRI is still refused. The namespace
there is the RDFLib serializer's own choice, reproduced on rdflib 6.3.2 and
7.6.0, and canonicalize_rdf_graph, which writes predicates in full,
serializes those graphs. No dependency grammar is patched locally.

Verification

• Source suite: 1,170 tests pass, 96.21% runtime coverage.
• Full suite at both dependency endpoints: 1,170 pass on rdflib 6.3.2 with
  pyoxigraph 0.5.4 and on rdflib 7.6.0 with pyoxigraph 0.5.11.
• Failing-before evidence: both example IRIs in subject, object and datatype
  position raised at the previous base and now serialize and reparse
  isomorphically through every Turtle-family alias, normal and degraded.
• Explicit before/after byte comparison over bound, unbound, shared-list and
  degraded graphs: RDF/XML, JSON-LD, N-Triples and pyoxigraph Turtle bytes
  unchanged; the Turtle-family delta is only the removed declarations.
• Ruff, mypy, reference validator and requirement checker pass offline.
• Exact-wheel suite: 1,168 pass with two expected source-only skips.
• The platform/Python CI matrix must pass before merge.

No public signature, dependency requirement or version changes.