Support custom PostgreSQL column types: RawText binds + select_as_text reads

[vsdudakov]

Problem

Custom column types (pgvector vector, inet, citext, ...) were unusable on PostgreSQL, even though the custom-fields guide advertises a pgvector VectorField as its flagship example:

• Writes: a text rendering returned from to_db binds with a declared text parameter type, and PostgreSQL has no implicit text → custom-type assignment cast — every ORM INSERT/UPDATE fails with SQLSTATE 42804. Downstream apps had to install a global CREATE CAST (text AS vector) WITH INOUT AS IMPLICIT via migration.
• Reads: the engine requests binary result values, which it cannot decode for a custom type — selecting the column raised unsupported PostgreSQL type OID, forcing .defer() workarounds on every read.

Fix

• yara_orm.RawText — a thin str subclass (same pattern as Array). On PostgreSQL it binds untyped: the engine declares OID 0 so the server infers the parameter type from context (the target column, a <=> operand), and sends the value with the text format code so the type's own input function parses it. Implemented as a Value::RawText variant in the Rust engine (pg_type → None, encode_format → Format::Text); every other backend treats it exactly like a plain string.
• Field.select_as_text — new class attribute; PostgreSQL SELECT/RETURNING projections read the column through CAST(col AS text) (a bare column cast keeps the column's result name, so decode plans are unaffected). Rendering goes through a new BaseDialect.select_column hook applied at every projection site: full-row selects, only()/defer(), select_related, values()/values_list(), m2m prefetch and RETURNING. Comparison and grouping references keep the column's real type.
• The custom-fields guide now documents both on a complete pgvector VectorField.

Tests

tests/test_raw_text_custom_types.py runs on every configured backend using a custom inet field kind — a built-in PostgreSQL type with no implicit text cast and no engine decoder, i.e. the exact pgvector failure shape without the extension dependency. Covers create/get/only(), filter/update/values_list(), bulk_create/save, select_related and m2m prefetch, plus unit tests for the dialect hook and the marker type.

Verified end-to-end against a real pgvector database: create, full-row get, update, bulk_create, values_list and a raw-SQL ORDER BY embedding <=> $1 KNN query with a RawText parameter (no ::vector cast, no CREATE CAST) all pass.

Full suite green (1221 sqlite under xdist + serial postgres modules), make lint clean, coverage unchanged vs main (identical missed-line set, all backend-specific code needing live mysql/mariadb/oracle/mssql servers).