Delegate reminder dispatch to PushNotificationService and log results

Implement the dispatch loop using Stream.fromIterable(candidates).asyncMap() with a throttle, or more practically, chunk the list into groups of 10 and use Future.wait() per chunk — this gives controlled concurrency without a full async stream overhead. The per-record log write (not batch) is critical for fault tolerance: if the Edge Function is killed at candidate 150 of 200, the first 150 are already logged as delivered and will be skipped on the next cron run via deduplication. Resolve all device tokens in a single SELECT WHERE user_id IN (...) before the loop and build a Map (mentorId → fcmToken) for O(1) lookups. When a mentor has no device token (never logged in on mobile), log a 'failed' record with failure_reason='no_device_token' — this is a legitimate failure mode that coordinators may need to act on.

The coordinator grouping: iterate the Map from task-010 in a separate loop from the mentor loop to avoid double-counting in DispatchSummary.

Unit tests (flutter_test): (1) Mock PushNotificationService returning success for all candidates; assert cert_notification_log receives one 'delivered' INSERT per candidate. (2) Mock PushNotificationService failing for candidates 2 and 4 in a list of 5; assert loop completes all 5, candidates 2 and 4 logged as 'failed', others as 'delivered', DispatchSummary has successCount=3 and failureCount=2. (3) Mock cert_notification_log INSERT failing after a successful push; assert no re-dispatch occurs and error is logged. (4) Assert coordinator payload dispatched once per coordinator, not once per underlying mentor candidate.

(5) Simulate 200 candidates with 10-concurrent dispatch; assert all complete within test timeout. Integration tests against local Supabase: verify cert_notification_log entries persist correctly; verify idempotency — re-running with the same candidates after deduplication results in zero new dispatches. Coverage target: 90% branch coverage on dispatch loop.