SCHEMA

This fix is a performance improvement.

1)

When a thread connects, reset of all per thread statistics
are now delayed until a statistic is actually collected.

This lazy initialization benefits workloads with very short lived sessions,
for which instrumentation is disabled.

2)

When a thread disconnect, the per thread statistics are aggregated
to a parent only for statistics that actually collected data.

This optimization benefits workloads with very short lived sessions,
for which instrumentation is disabled.

3)

For the statement instrumentation,
reset of an individual event_name statistic is also now delayed
until a statistic is actually collected.

This benefit all workloads, because all workloads only contain a few
types of statements (SELECT, INSERT, UPDATE, DELETE, ...),
from the very long list of statements supported in MySQL.

Only statements for event names actually executed are aggregated on
disconnect.

4)

The memory footprint of internal memory buffers is reduced,
by removing some attributes reserved for future use,
that were in fact not used.

In particular, statistics for mutexes, rwlocks and conditions
now need less memory.