refactor(codec): factor shared schema/signature helpers out of UDF encoders

Scalar, aggregate, and window UDF encode/decode bodies in codec.rs each
contained the same five blocks:

* a `match TypeSignature::Exact { ... } else err` to extract input
  dtypes,
* an `arg_{i}` input-field synthesis with a verbatim 6-line comment,
* a `Schema::new(vec![Field::new(...)])` for the return type,
* a `schema_from_ipc_bytes(...).first().ok_or_else(...)` decode of the
  single-field return blob, and
* a `parse_volatility(&volatility_str).map_err(...)` round-trip.

Six near-identical bodies meant the same comment text lived in three
places, and each `map_err(|e| PyValueError::new_err(format!("{e}")))`
chain appeared a handful of times per body. Extract:

* `signature_input_dtypes(sig, kind)` — `Signature::Exact` extraction
  with a flavor-tagged error.
* `build_input_schema_bytes(&[DataType])` — synth `arg_{i}` fields and
  write IPC. Carries the comment explaining why field metadata is
  discarded on decode.
* `build_single_field_schema_bytes(&Field)` /
  `build_schema_bytes(Vec<Field>)` — IPC writer wrappers.
* `read_input_dtypes(&[u8])` /
  `read_single_return_field(&[u8], kind)` — decode side.
* `arrow_to_py_err(ArrowError) -> PyErr` and `parse_volatility_str` to
  collapse the repeated `map_err` chains.

Wire format unchanged. Six encode/decode bodies collectively shrink
from ~300 to ~140 LOC and stop carrying triplicated comment text that
would otherwise drift independently.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: timsaucer