DurationStatsTracker is highly inefficient

DurationStatsTracker is calculating running average in the fast path, incurring at least 4 volatile reads, 2 double-precision multiplies, 1 double-precision division and two compare-and-swaps. Optionally this includes another 2 CASes and 2 volatile writes. Reading the stats afterwards is a breeze, though.

This is obviously biased in the wrong direction, so it needs fixing.