Add json_union_to_text UDF to flatten a JSON union to JSON text

[adriangb]

json_get returns a heterogeneous JSON union (JSON_UNION_DATA_TYPE). Consumers that can't represent an Arrow Union — notably the Parquet writer, whose arrow_to_parquet_schema panics on unions (unimplemented!("See ARROW-8817.")) — have no way to materialize such a column. None of the existing UDFs help: they all take (json_string, path…) and parse source text rather than consuming a union.

This adds json_union_to_text(json_union), which flattens the union (via the public JsonUnionEncoder) into canonical JSON text as Utf8View:

• scalars render as true / 42 / 1.5
• strings are JSON-quoted and escaped
• array/object arms (already raw JSON text) pass through
• a JSON null arm becomes SQL NULL

Signature::exact([JSON_UNION_DATA_TYPE]) constrains the input to the JSON union, so any other argument type is a planning error. Registered in register_all and exported under functions:: / udfs::, matching the other functions.

Tested: cargo test --lib json_union_to_text (covers every union arm incl. escaping), cargo clippy --all-targets (pedantic), cargo fmt --check.

[cetra3]

Can we copy what happens with arrow-rs and use serde json to do this? I worry deeply by hand rolling our own encoder we are going to hit a number of edge cases & bugs. serde_json is already a dependency of datafusion/arrow:

https://github.com/apache/arrow-rs/blob/58.3.0/arrow-json/src/writer/encoder.rs

[adriangb]

Done — switched the scalar arms to encode via serde_json::to_writer (added serde_json as a direct dep, since it's already in the tree via arrow). The array/object arms still pass their raw JSON text through verbatim, since those are already-serialized JSON substrings (re-encoding them through a JSON string encoder would double-quote them). Dropped the hand-rolled escaper entirely and extended the test to cover quote/newline/control-char escaping.

[cetra3]

Approved but I would like to use serde_json to do the encoding

[adriangb]

Also marked the output field as JSON: the UDF now implements return_field_from_args to return a Utf8View field tagged with the canonical Arrow JSON extension via json_field_metadata() (ARROW:extension:name = arrow.json), matching json_get_json/json_get_array. Added a test asserting the extension name on the output field.