* fix selected_gpus bug for local_cluster

* fix bugs about single machine multicards training for local_cluster

* fix python -m config.yaml -b backend.yaml

* re add selected_gpus_num config for single train mode

* fix bug for submit for paddlecloud k8s
Co-authored-by: Nwuzhihua <35824027+fuyinno4@users.noreply.github.com>