Ceph

The version in master is problematic because the log message can race with a client update messing up the log segments persisted on a replica. Additionally, the testing is pretty rudimentary: it doesn't run under thrashing.
1) Create more general mechanism for merging a set of log entries into current actingbackfill replicas which tolerates concurrent client IO
2) Adapt lost-unfound commands to use 1)
3) Use 1) to create a debug interface to inject arbitrary json entries -- that will let us create bogus unfound objects on command.
4) Use 3 and 2 to test the lost-unfound code during thrashing.