Reenable WAL fsync in the master

After fixing KUDU-1125, I ran a quick experiment on an internal cluster to test the performance impact of enabling log_force_fsync_all in the master. I observed that a tablet report with 308 tablet updates took 157 ms. I think this is good enough for us to restore the default value to true for Kudu masters so that we'll have real metadata durability in the event of a crash.

I don't know whether we should do this for all master deployments or just for single-master ones. Certainly the latter would benefit most, but it might be annoying to detect the difference.