The bug was that if mysql.slave_gtid_pos was missing, operations on variables
gtid_slave_pos, gtid_binlog_pos, and gtid_current_pos would fail, and continue
to fail even after the table was fixed, until server restart.

Now setting the variables retry loading the table, succeeding if it has been
restored. And querying the variables when the table is not there acts as if
the table was there and was empty.

Also, attempt to fix a race in the rpl.rpl_gtid_basic test case.