0.9.4

0.9.4 (2017-03-07)

• Strictly enforce the cluster move state machine
• Don't start moving new blocks while existing moves are in progress
  (fix bug where the same destination block could be reused)
• Hold a lock to exclude flush while updating references to ensure
  reference updates hit the disk before the move is considered complete
• Simplify allocator by always adding blocks to the Roots set before
  returning. The caller must transfer them somewhere else.
• Simplify the cluster moving API by combining get_moves with
  start_moves, so it's not possible to block and affect the moves
  which can legally be started
• When detecting a duplicate reference or hitting an I/O error, log
  analysis of the internal state
• Check for move cancellation before copying a block to avoid accidentally
  copying a block which is now outside the file
• Avoid adding a cluster to the Junk set twice during a reference update
• Add lots of assertions