Scanning backwards can be really slow if replication was paused for significant time.

Theory: scanning backwards is horribly slow because every 4k have to seek

Solution:

• scan bacwards as we do now for 5 seconds.
• if the starting point is not found, find the extent which the object should be in by looking at first object in each extent
• then scan forwards in that extent - which should be very fast

If we can do this simply - might be worth backporting to 1.2.x