Crash occurred because server was not ready to execucte delete of
rows in ndb_binlog_index table before server is properly started.
Changed behavior when binlog purge is requested before the server
is properly started to delay the actual purge of ndb_binlog_index.
By saving purge requests in a queue they can be done by the binlog_thread
at startup and when the server is properly started.
Added new ndb_binlog_purge_startup test that forces purging of all
binlog files at startup.