wal: handle Aurora's pg_ls_waldir() unsupported error gracefully

[megative]

Summary

Fixes #1309.

Amazon Aurora PostgreSQL rejects pg_ls_waldir() with feature_not_supported (Postgres error class 0A). The wal collector currently treats this as fatal, aborting every scrape on Aurora and producing no WAL metrics.

This PR detects the Aurora-specific error fingerprint (class 0A + "Aurora" in the message) and short-circuits cleanly: returns nil from Update without emitting metrics on Aurora. Other database errors still propagate normally.

Design notes

• Non-emission over NaN. Per @sysadmind's feedback on fix(collector): handle Aurora's unsupported pg_last_xact_replay_times… #1274, the convention here is to omit metrics when the underlying value is unavailable, rather than emit a NaN sentinel.
• Narrow error fingerprint. The check requires both Postgres error class 0A AND "Aurora" substring in the message, so non-Aurora users see no behavior change.
• Independent of fix(collector): handle Aurora's unsupported pg_last_xact_replay_times… #1274. Touches different files (pg_wal.go / pg_wal_test.go); no overlap with @dannotripp's pg_replication fix.

Test plan

• go build ./..., go vet ./..., go test ./... clean.
• New test TestPgWALCollectorAurora verifies the Aurora fallback: collector returns nil and emits zero metrics when the query returns the Aurora-specific error.
• Vanilla PostgreSQL 17 unaffected — pg_wal_segments / pg_wal_size still emit normally.
• Live-tested against Aurora PostgreSQL 17.7.0 — wal collector still reports success=1 and continues to emit pg_wal_segments (and pg_wal_size_bytes when non-NULL).