fix: Rebuild projections once per context after the registry changed

Every ephemeral OfflineSyncDatabase shares one context but initializes a
recorder of its own, and each rebuilt the projections of every space
when the context's first initialize had changed the registry: on every
session and every merge, until the process stopped (upstream #150). A
new device, or one whose app changed its synced tables, spent about 3 s
of every inbound merge on it with a library of 8,000 cards, a server
after such a migration as much on every request.

The context now runs the rebuild once (Once): wrappers initializing
while it runs await the same run, later ones skip it, and a failed run
is tried again by the next.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

packages/serverpod_offline_sync/test/utils/once_test.dart

@@ -0,0 +1,47 @@
+import 'dart:async';
+
+import 'package:serverpod_offline_sync/src/utils/once.dart';
+import 'package:test/test.dart';
+
+void main() {
+  group('Given an action run once,', () {
+    late Once once;
+    late int runs;
+    late Completer<void> gate;
+
+    setUp(() {
+      once = Once();
+      runs = 0;
+      gate = Completer<void>();
+    });
+
+    Future<void> action() async {
+      runs++;
+      await gate.future;
+    }
+
+    test('when asked while it runs and after, then it runs once and every '
+        'caller awaits that run.', () async {
+      final first = once.run(action);
+      final meanwhile = once.run(action);
+      gate.complete();
+      await Future.wait([first, meanwhile]);
+      await once.run(action);
+
+      expect(runs, 1);
+    });
+
+    test('when its run fails, then the next caller runs it again.', () async {
+      final failing = once.run(() async {
+        runs++;
+        throw StateError('failed');
+      });
+      await expectLater(failing, throwsStateError);
+
+      gate.complete();
+      await once.run(action);
+
+      expect(runs, 2);
+    });
+  });
+}