connection threads from cleaning up.

Analysis
=========
Issue here is, delay in connection cleanup for which global
connection counter is already decremented, makes room for
new connections. Hence more than expected connections are
observed in the server.

Connections running statement "SHOW PROCESSLIST" or "SELECT
on INFORMATION_SCHEMA.PROCESSLIST" acquires mutex
LOCK_thd_remove for reading information of all the connections
in server. Connections in cleanup phase, acquires mutex to
remove thread from global thread list. Many such concurrent
connections increases contention on mutex LOCK_thd_remove.

In connection cleanup phase, connection count is decreased
first and then the thd is removed from global thd list. This
order makes new connection (above max_connections) possible
while existing connections removal is still pending because
of mutex LOCK_thd_remove being held by show processlist.

Fix:
=====
In connection clean phase, thd is removed from the global
thd list first and then global connection count is
decremented. Added code to wait for connection_count be
be zero during shutdown to ensure connection threads are
done with their task.

Because of this change, tests which verifies PROCESSLIST and
then status variable "threads_connected" might fail on slower
machines. As part of this patch, changed test cases to wait
until "threads_connected" is set to expected value instead
of PROCESSLIST count.