Make driver pod restart policy configurable

[liyinan926]

To support driver high availability, particularly in Spark Streaming use cases, we need to allow applications to specify the driver pod restart policy, instead of hard coding Never.

[felixcheung]

I think this is very important for job reliability in general (esp. in the cloud?) and not just for streaming

[liyinan926]

I agree.

[ChenLingPeng]

What do we expect to happen after driver restart, should/can we reuse the pre-created executors?