Fix bug in stats_event_benchmark

Now, the number of elements in each simulated stats_event is correctly
sampled from a uniform distribution.

Rerunning the benchmarking tests indicates that truncating the buffer
increases the number of cycles require to log an event by only 1.6%.

Test: m libstatssocket_benchmark
Test: bit libstatssocket_benchmark:*
Change-Id: I8d5c1734861038beec8eeb7ce1f84a24bde25a75