Add builtin varchar array codec

[anton-oai]

Summary

Add a builtin binary codec for PostgreSQL varchar[] / pg_catalog._varchar.

asyncpg already has builtin support for scalar varchar and for text[], but not for varchar[]. As a result, queries that return or bind varchar[] require asyncpg to run its recursive typeinfo introspection query on each new connection.

This change registers varchar[] as a builtin array codec. It reuses text element encode/decode behavior, while encoding arrays with VARCHAROID as the element OID so bound varchar[] parameters match PostgreSQL's expected array element type.

Motivation

Without an explicit builtin codec for varchar[], asyncpg runs its typeinfo introspection query for every new connection that sees this type.

In isolation, the cost of one introspection query is low in our measurements, around ~200us when a single query is executed. However, during reconnect storm events with thousands of concurrent reconnects, this introspection query becomes a major CPU hit and can materially degrade overall database performance.

For schemas with many varchar[] columns, especially on large tables where changing column types is not desirable, supporting this native PostgreSQL type out of the box is more practical than requiring applications to avoid varchar[] or add query-level casts.

Test Plan

Added test_varchar_array_does_not_introspect, which uses a custom connection class that counts _introspect_types() calls and verifies that both parameter and result varchar[] paths work without triggering type introspection.

Locally run:

PYTHONPATH=. .venv/bin/python -m pytest tests/test_introspection.py -k varchar_array_does_not_introspect -q