ShardingTest.stop must join top-level chunk migrations

The SharingTest stop method is calling the check orphans hook without previously ensuring that top chunk migrations triggered by the chunk splitter have finished.

As a result, the following race condition can happen when the balancer is active:

• A migration stops on the receiving side, waiting for an overlapping range deletion to end
• The test finishes, ShardingTest.stop is called to shutdown the cluster
• The check orphans hook waits for range deletions to drain
• Overlapping range deletion finishes:
  • The hook finishes waiting
  • A new pending range deletion is enqueued and the clone of documents from the donor begins
• The hook misidentifies the just cloned documents as orphans