Task executors that support specific roles are restarted when they fail

[zuston]

Why

Now TonY introduces the Sidecar Tensorboard, but sometimes it will fail due to hardware problems and unstable HDFS. But for users, it's better to unconscious restart it.

So we need to introduce the general mechanism to meet above requirements

[zuston]

@oliverhu Please check it.

Further more, maybe elastic training also need above feature.

[oliverhu]

Sounds good to me