Implement rollback logic on read state mutation failure

In the catch block of both `markAsRead` and `markAllRead`, retrieve the stored snapshot from `_snapshot[id]` (or the bulk snapshot map) and emit `ReadStateRollback(ids: [...], restoredStates: {...})` followed by `ReadStateFailure(message: _humanReadableError(e))`. Implement `_humanReadableError` as a switch on exception type: `SocketException` → 'No internet connection', `TimeoutException` → 'Request timed out, please retry', `PostgrestException` with code 42501 → 'You do not have permission to update this notification', default → 'Something went wrong, please try again'. Always remove IDs from `_inFlight` in a `finally` block to prevent permanent stuck state. Clear snapshot entries after use (both success and failure) to avoid memory growth.

For idempotency: check if notification is already in un-read state before applying rollback — if already rolled back, skip silently.

Unit tests (flutter_test) using mock repository that throws configurable exceptions. Test scenarios: (1) network error on single markAsRead triggers rollback of that item only, (2) timeout on markAllRead triggers bulk rollback of all optimistically updated items, (3) RLS rejection produces 'permission denied' human-readable message, (4) retry after rollback produces fresh optimistic update (verify snapshot is cleared and re-captured), (5) rollback with missing snapshot emits failure without throwing, (6) rollback is idempotent (call rollback handler twice, assert state unchanged on second call). Use fake_async to simulate timeouts. Verify in-flight set is empty after both success and failure paths.