* make job run task in a separate process

* fix typo

* add more debug info in xparl client

* refine control flow of different processes in xparl job

* refine control flow of different processes in xparl job

* remove tsinghua source

* remove tsinghua source

* remove unnecessary logic

* fix typo

* refine comments and some logic

* fix bug, `decay=0` means totally synchronize weights of source model to target model