cancel_sync is blocked when synchronisation is blocked by a memory alarm on the receiving node

[hairyhum]

To reproduce:

1. Create a lazy mirrored queue
2. Publish a lot of big messages (e.g. 4000 1MB messages)
3. Restart a slave with watermark around 300M
4. Start syncing
5. See the queue stuck syncing
6. Try to cancel sync

Affected version: master, 3.7

A syncing queue should somehow detect being stuck and at least support cancelation and policy changes. Maybe purge as well, to give user a way to exit the stuck state other than restarting nodes.