feat(postgresql): add analyzerv2 experiment for database-only analysis

[kyleconroy]

Summary

This PR introduces a database-only analyzer mode that uses the database connection for query analysis instead of sqlc's internal catalog. This provides more accurate type inference by leveraging the actual database schema.
Configuration

The feature is enabled by setting analyzer.database: only and requires the analyzerv2 experiment:

version: "2"
sql:

• engine: postgresql # or sqlite
  schema: "schema.sql"
  queries: "query.sql"
  database:
  managed: true # or uri: "postgres://..."
  analyzer:
  database: "only"
  gen:
  go:
  package: "db"
  out: "db"

Run with SQLCEXPERIMENT=analyzerv2 to enable:

SQLCEXPERIMENT=analyzerv2 sqlc generate

Key Features

Database-backed type inference: Uses PREPARE statements to get accurate column types
Star expansion: Expands SELECT * and RETURNING * using the database
Syntax validation: Schema files are parsed to validate SQL syntax and catch errors early
PostgreSQL and SQLite support: Works with both database engines
Experiment-gated: Requires SQLCEXPERIMENT=analyzerv2 to activate

How It Works

Schema migrations are parsed for syntax validation (errors are reported)
Schema contents are passed to the database connection
The internal catalog remains empty - database is the source of truth
Queries are analyzed by preparing them against the database
Star expressions are expanded using column names from prepared statements

Changes

Added analyzer.database config option with true, false, and "only" values
Added analyzerv2 experiment flag
Extended Analyzer interface with EnsureConn and GetColumnNames methods
Added SQLite analyzer support for database-only mode
Refactored compiler to use unified analyzer interface

Test plan

PostgreSQL tests pass (accurate_cte, accurate_enum, accurate_star_expansion)
SQLite tests pass (accurate_sqlite)
Schema syntax errors are properly reported
Feature only activates with experiment flag