Release pgpq 0.12.0 and arrow-json 0.10.0

[adriangb]

Merging this publishes. detect-release sees a version the registry does not have and the
release jobs push pgpq to crates.io and PyPI, and arrow-json to PyPI.

• pgpq 0.11.1 → 0.12.0 (core/Cargo.toml, py/Cargo.toml, py/pyproject.toml)
• arrow-json 0.9.0 → 0.10.0 (json/pyproject.toml; json/Cargo.toml tracks the workspace
  version as it always has)
• Cargo.lock and uv.lock regenerated

What's in 0.12.0

The breaking changes were deliberately batched so they would cost one release rather than several.

Breaking

• arrow 56 → 59 and the public API that moves with it (Update dependencies: arrow 59, pyo3 0.29, thiserror 2, Dependabot #82).
• write_header returns Result — the encoder state machine reports misuse instead of
  aborting the host process (Refactor encoders to generics; return errors instead of panicking on misuse #89).
• Nested composites require their OID (Composite/array OID correctness: dummy 16385 collides with real user OIDs; Json.oid() returns jsonb's 3802 #96/Take composite OIDs from the caller instead of guessing 16385 #102). PostgresType::UserDefined carries
  oid: Option<u32>; supply it with ArrowToPostgresBinaryEncoder::with_composite_oids, keyed by
  the type name the DDL creates, and use composite_type_names() to learn which names to look up.
  This replaces a hard-coded 16385 that silently collided with real user OIDs on any database
  that had ever created a user-defined type.
• Int8 can no longer be written as Char (Int8 Char output type is broken: INT2 payload declared as bpchar text, rejected by Postgres on every value #95/Refuse Int8 -> Char instead of emitting output Postgres rejects #101). It declared bpchar for a two-byte
  INT2 payload, so Postgres rejected every value; it is refused at build time now.
• Composite DDL uses the Arrow field names rather than positional f0, f1 (Struct support polish: carry real field names into CREATE TYPE DDL; export StructEncoderBuilder to typed Python #97/Carry Arrow field names into composite DDL; export StructEncoderBuilder to typed Python #99), and
  every identifier the DDL emits is quoted.

Fixes

• Json.oid() returned jsonb's 3802 instead of json's 114, so a json[] column was rejected by
  the server (Composite/array OID correctness: dummy 16385 collides with real user OIDs; Json.oid() returns jsonb's 3802 #96/Json.oid() must be json's 114, not jsonb's 3802 #100).
• FixedSizeBinary/FixedSizeList encoders implemented; struct-with-list-field panic fixed (Implement FixedSizeBinary/FixedSizeList encoders; fix struct-with-list-field panic #94).

Performance

• Encoding is roughly 2× faster on the NYC taxi benchmark, wire format byte-identical (Encoder hot-path optimizations (measured) #98).

Also

• StructEncoderBuilder is exported to typed Python (Struct support polish: carry real field names into CREATE TYPE DDL; export StructEncoderBuilder to typed Python #97/Carry Arrow field names into composite DDL; export StructEncoderBuilder to typed Python #99).
• Generative-suite polish: sign-of-zero fidelity, unified float equality, wider fuzz coverage,
  and the snapshot corpus blind spots closed (Generative suite polish: sign-of-zero fidelity, unified float equality, fuzz coverage, snapshot corpus gaps #93).

The README's data type table is refreshed in the same commit — it still claimed structs were
unsupported, still offered CHAR for Int8, and omitted LargeUtf8, Utf8View, LargeBinary,
FixedSizeBinary, LargeList and FixedSizeList. It now also documents the composite-OID flow.

Verification

cargo test --locked --all-features (290 tests) and the full pytest suite (25, including the
9 Postgres-backed ones) green locally against a real server, plus cargo fmt, clippy,
uv lock --check and pre-commit.