Our opLatencies (i.e. Top) tracking is done in the destructor of AutoStatsTracker.

This is problematic for several reasons:

• This excludes processing time that occurs before AutoStatsTracker is constructed and after it is destructed
• This is inconsistent with the latencies we report for slow queries and profiling
• This makes an extra call to read the clock, that would not be needed if we reported the Top stats in the same place we call completeAndLogOperation().
• Each command implementation is responsible for correctly using AutoStatsTracker, which means there is no enforced consistency with latency measurements.