Assert directory fragmentation is occuring during stress tests
Currently we enable fragmentation and set a low (100) frag size limit, but nothing actually validates that there is any splitting going on, especially during thrashing runs where we care the most.
We have performance counters on the MDS that will tell us if splitting has occurred, but they will reset when MDSs restart.
We could add something to the thrasher task that captures MDS counters before shutting down, and accumulates values. We don't necessarily need this on every task, just something in there to check that splitting is really happening during at least some stress testing.