(1) set fleet_send_batch_num a default value according to trainer num， the previous 80000 is fixed，if trainer num is much less or larger than 100，global shuffle may have timeout error.

(2) fix load one table bug, add barrier