feat: per-variant constructors for data enums in the dynamic backends

[tobocop2]

Closes #147.

What

Generate a per-variant factory constructor for every data-carrying variant of an internally-tagged data enum, in the dynamic backends. The statically-typed backends already derive per-variant constructors from the variant structure; this brings the dynamic backends to the same model instead of leaving them on stringly-typed new(type=...) / raw maps.

Why

For the same enum, construction diverged purely by target. Take EmbeddingModelType (Preset { name }, Custom { model_id, dimensions }, Llm { llm }, Plugin { name }):

Static backends — already type-safe, the variant name is the constructor:

new EmbeddingModelType.Preset("balanced")          // C# / Java

.preset(name: "balanced")                          // Swift / Dart

Dynamic backends, before — stringly-typed, no per-variant constructor:

EmbeddingModelType(type="preset", name="balanced")   # pyo3 — magic "preset" string

{ type: 'preset', name: 'balanced' }                 # magnus — raw Hash

EmbeddingConfig::from_json(json_encode(["model" => ["type" => "preset", "name" => "balanced"]]))   # php — empty class, hand-built JSON

"{\"model\":{\"name\":\"balanced\",\"type\":\"preset\"}}"   # rustler — literal JSON string

Dynamic backends, after — per-variant constructors, parity with the static backends:

EmbeddingModelType.preset("balanced")
EmbeddingModelType.custom(model_id="BAAI/bge-small-en-v1.5", dimensions=384)
EmbeddingModelType.llm(cfg)
EmbeddingModelType.plugin("my-backend")

EmbeddingModelType.preset("balanced")

EmbeddingModelType::preset("balanced");

EmbeddingModelType.preset("balanced")

Type-safe (no magic "preset" string), discoverable via autocomplete, and complete — it covers Custom / Llm / Plugin, which the data-variant bare-string shorthand never addressed.

string_shorthand was merged in #133 to address #135, but it isn't a good design and is reverted here. A bare string is stringly-typed, only ever worked for the single-field Preset case, and only on pyo3 and magnus — it doesn't extend to Custom / Llm / Plugin or to the other backends. This PR addresses the same goal properly: every variant gets a typed per-variant constructor, the same way across all backends. #135 is superseded.

The unit-variant string handling from #132 is preserved and unaffected — a fieldless variant name ("disabled") is a fine string, and per-variant constructors don't replace it.

How

• Derive one constructor per data struct variant from the variant's fields; body constructs the core variant directly (Self { inner: <core>::<Variant> { field, .. } }), reusing the existing param / let-binding / conversion machinery.
• Each constructor collides with the variant accessor of the same name, so it uses the _factory_<name> Rust ident + the host-facing <name> (pyo3 #[pyo3(name=...)], etc.), per backend.
• A hand-written impl constructor of the same name suppresses synthesis (consumer wins). Unit, tuple, and binding_excluded variants are skipped.
• Static backends already conform — audited, no change.

Verification

Regenerated kreuzberg against this branch and checked every impacted binding:

• All five backends emit the per-variant constructors on the real EmbeddingModelType / RerankerModelType (and other data enums) — EmbeddingModelType.preset(...) / .custom(...) / .llm(...) / .plugin(...) in Python, EmbeddingModelType.preset(...) in Ruby, ::preset(...) in PHP, EmbeddingModelType$preset(...) in R, EmbeddingModelType.preset(...) in Elixir.
• cargo check passes on the regenerated kreuzberg-py, kreuzberg-php, and kreuzberg-node. This surfaced (and the fix commits resolve) several real-world field-type cases the neutral unit fixtures missed: variants with sanitized / binding-excluded fields are skipped (they can't be built from the binding), promoted-optional params unwrap to non-optional core fields, return-only DTOs get a generated From impl, and field conversions are inlined so no non-re-exported core type path is named.
• pyo3/magnus: enum constructors reject bare variant strings for internally-tagged enums #132 (the unit-variant bare-string {"<tag>": s} wrap) is preserved across pyo3 and magnus; string_shorthand is fully removed with no dangling references.

The shared collect_variant_constructors (variant selection) and variant_field_init (field conversion) are the single source of truth for the wrapper-convert backends; magnus builds the binding enum directly; rustler emits pure-Elixir constructors. Whole-crate cargo test, clippy -D warnings, and fmt are clean.

[Goldziher]

Merged into main. Reviewed end-to-end: cargo fmt --check, clippy --all-targets, and the full test suite pass on the merge result.

Two trivial adjustments during merge (neither touches the PR's backend logic):

• Resolved a CHANGELOG.md conflict by keeping both Unreleased sections (the per-variant-constructor entries plus a concurrent e2e change), ordered Added → Changed → Removed.
• The base had an unrelated cargo fmt nit and a few project-name mentions in e2e delegation comments that tripped the cli_no_project_special_casing guard; both fixed on main (separate commits), unrelated to this PR.

Thanks @tobocop2 — the draft body was stale; the work was complete across pyo3/magnus/php/extendr/rustler with the string_shorthand removal.