HAMMER 61E/Many: Stabilization, Performance

HAMMER 61E/Many: Stabilization, Performance

* The UNDO FIFO could get over-full, causing an assertion.  This was because
  certain load tests could cuase directories to end up with tens of
  thousands of directory entries and all of them were being flushed in
  a single flush group.

  Rewrite the flush group infrastructure to fix the issue.  Instead of
  having a two-stage flush we now have flush compartmentalization with
  the new hammer_flush_group structure.  Flushes can thus be broken down
  into transactions that don't blow out the UNDO FIFO.

* Improve flush performance by quite a bit.  The new infrastructure
  generates fewer B-Tree conflicts and appears to flush the work more
  smoothly, probably owing to the fact that the upper size of each
  flush group is now bounded.