Rewrite index stats loop so that required objects, Ndb
in particular, can be checked in each processing slice.
On system restart (initial or gci gap) binlog thread sets
ndb_index_stat_restart flag.  Stats thread when waking up
again then re-creates or re-discovers required objects.

This however leaves a gap where clients can make queries
before stats thread is ready (error 4721).

Second part of the fix is to prevent clients before
stats thread is ready.  This also means that mysqld with
index stats enabled will not start before stats thread
is ready.  Previously this was more sloppy, any problems
probably masked by delays etc.

Clients discover stats thread state from the "allow" field
in Ndb_index_stat_status.  This needs to be checked on
cluster (initial) restart and also is index stats enable
has been turned off and on again.  Known cases in MTR
have been fixed.