feat: implement instance backup and restore system

[hippoley]

Instance Backup System

Implements the complete backup system for OpenClaw instances, filling in the empty backups and backup_schedules DB schema that was already defined in 001_init_schema.sql.

Part 1: Manual Backup & Restore

• BackupRepository: CRUD operations for the backups table
• BackupService: full lifecycle — create, list, get, delete, restore
• K8s Job-based backup mechanism: tar.gz archiving of HostPath instance data via busybox Jobs
• Async job monitoring with race-condition-safe status updates (markBackupCompleted/markBackupFailed re-read DB state before writing)
• TTLSecondsAfterFinished for automatic K8s Job cleanup
• BackupHandler: 5 HTTP endpoints (create, list, get, delete, restore)
• 13 unit tests covering service logic, validation, ownership checks, and edge cases

Part 2: Scheduled Backups & Expiry Cleanup

• BackupScheduleRepository: CRUD + ListAllActive() for the backup_schedules table
• BackupScheduler: background loop (60s tick interval) with a minimal cron expression parser
  • Supports @hourly, @daily, @weekly, @monthly presets and standard 5-field cron (minute hour dom month dow)
  • Field tokens: *, numbers, ranges (1-5), steps (*/6, 1-5/2), lists (1,3,5)
  • All times evaluated in UTC, consistent with K8s CronJob behavior
• CreateScheduledBackup: system-actor backup creation (no user-ownership check) with auto-computed expires_at
• BackupScheduleHandler: 4 HTTP endpoints for schedule CRUD with cron expression and retention_days validation

Risk Mitigations (built-in)

• Idempotency guard: 90-second gap check prevents double-fire within the same cron window
• Expiry cleanup safety: only soft-deletes backups with status = "completed" AND expires_at < now() — never touches creating status
• Retention minimum: retention_days >= 1 enforced at API level
• Cron parse failure: gracefully returns false + logs warning, never panics
• Panic recovery: tick() has defer recover() so the scheduler goroutine survives unexpected panics
• Stop() double-call safety: sync.Once protects close(stopChan) from double-close panic
• Tick overlap prevention: sync.Mutex.TryLock() skips a tick if the previous one is still running

Zero Breaking Changes

• All modifications to existing files are pure additions (no existing function bodies modified)
• No route conflicts with existing endpoints
• DB schema already exists — no new migrations needed
• Scheduler integrates with existing graceful shutdown flow

Test Coverage

• 23 total tests (13 Part 1 + 10 Part 2)
• Part 2 tests: cron presets, standard expressions, ranges/steps/lists, invalid expressions, cron validation, scheduler idempotency, cleanup safety, invalid cron skip, double-stop safety, panic recovery
• Full regression: all existing tests pass