refactor: let rdflib decide where collection syntax is safe

[jdsika]

Why

deterministic_turtle carried a TurtleSerializer subclass that re-decided where ( … )
collection syntax may be used, rejecting any chain with a cell more than one statement points
into. That decision belongs to rdflib — and it now makes it, with the same condition, in both of
its Turtle serializers (RDFLib/rdflib#3504).

Two code paths claiming the same thing is one too many. The subclass goes; the first render uses
rdflib's serializer directly.

What stays

The part that is this library's own job: render, check the text round-trips, and fall back to
explicit rdf:first/rdf:rest statements when it does not. That check is what makes the
guarantee hold, and it holds whichever way rdflib decides.

Two consequences worth naming:

• No rdflib floor bump. The fix is merged upstream but unreleased (24 commits ahead of 7.6.0),
  so a floor could not name it anyway. Because correctness comes from the check rather than from
  the serializer's choice, the library is correct on any rdflib>=6.0 and picks up nicer output
  automatically when a fixed rdflib ships — no second change needed.
• A future rdflib that inlines more aggressively cannot silently break the canonical form.

Verified inert, not assumed

• The deleted subclass and rdflib's serializer emit byte-identical text for all four ASAM OpenX
  artifacts this is used on — opendrive and openscenario, OWL and SHACL (8,089 / 5,476 /
  15,203 / 11,040 triples). Identical text means the same round-trip verdict, the same fallback
  branch, and the same final bytes: no committed artifact changes, nothing needs regenerating.
• 188 tests pass. The six arrangements in SHARING_CASES now exercise the fallback rather than
  the subclass, so the sharing cases are still covered end to end.
• Readability unchanged: a private list still gets ( … ), a shared list still falls back, and a
  graph mixing both keeps ( … ).

Note on those artifacts

On the real ASAM graphs the deleted subclass was already a no-op — rdflib's inline form does not
round-trip them (openscenario gains 74 triples, opendrive 152, in the sh:in enumeration lists),
and the subclass produced exactly the same broken text, so the collection-free fallback is what
has been producing those files all along. Their correctness rests on the round-trip check, which
is precisely the part this PR keeps.

Worth a separate look, and deliberately not bundled here: those graphs round-trip only via the
fallback even though every chain containing a shared cell is rejected, which suggests a second
inlining path that neither condition covers. If that reproduces standalone it is a further rdflib
report, not something to work around here.

Not bundled

The module and test docstrings still carry provenance references (linkml#3295, issue #1) that
sit oddly with "docs state the current design". Left alone — unrelated to this change.