Fix DuckLake query startup with an empty catalog

[viggy28]

Closes #89.

What changed

• Detect whether the attached DuckLake catalog contains the PostgreSQL public schema before setting the query search path.
• Use streambed.main while the catalog is empty.
• Prefer streambed.public once the writer creates it.
• Keep the existing fresh, read-only attachment per client query, so the new schema becomes visible without restarting Streambed.
• Add regression coverage for both DuckDB-backed and SQLite-backed DuckLake catalogs.

Why

DuckDB rejects a search path containing a nonexistent catalog schema. A fresh DuckLake catalog has main, while Streambed creates public only after the first PostgreSQL relation is flushed. The previous unconditional search path made sync --target-format=ducklake --query-addr=... fail during startup.

Validation

go test ./internal/server -count=1
go test ./internal/... ./config/... -count=1

Both pass. The regression test starts the query server before public exists, flushes the first table through the writer, then verifies an unqualified query resolves it without recreating the server.