This work is split off from WT-4298.  We'll need to modify our latency workload used by workgen in bench/wtperf/runners/checkpoint-latency.wtperf :

• turn on snappy compression
• turn off logging on tables, leave logging on for the log-like file, if this is easy. This matches MongoDB. Alex doesn't think this will make much difference in driving our tuning decisions.
• MongoDB does in memory logging, with an explicit log_flush 10 times/sec. It would be good if we can do this in workgen. Again, only if easy.
• Higher granularity in monitor. Right now, we are gathering latency statistics once a second, I think more often may give us better insights (are spikes clustered within the checkpoint? happening at the beginning, or end of checkpointing?)