Ceph

this exercises some additional code paths. shouldn't be too frequent by default.

the 'wrongly marked down' thing presents as an ERR in the cluster log, so it would be nice if we could mask that particular error as well? or not, hmm!

(maybe the frequency of all the thrasher events should be tunable in the yaml, actually? so you can adjust the probabilities for different tests? then i could use it, say, only mark osds down with low frequency, or later use it to hammer on just the in/out code, or whatever...)