Improve RMNodeStarter and scheduler server network interface binding strategy

[fviale]

When several network interfaces are available on the host, both scheduler/rm and nodes often bind to the wrong one, making network configuration nearly unavoidable.

The purpose of this task to improve the binding strategy using the following mechanism:

• The scheduler server will bind preferably to the interface which has an Internet connection (if no internet connexion is available, it will fallback on the standard binding mechanism).
• The node will preferably bind to an interface which can connect to the provided Resource Manager url (in case of pamr, it will check the configured router and port).

[tobwiens]

What is the standard binding mechanism?

[fviale]

The standard binding mechanism is to try to bind to a public API address preferably, and if no public address is available (which is most often the case), to bind to the first private IP available (which can be anything).

[fviale]

Some improvements have been made to the binding strategy, but there are still cases where the server will bind for example to a docker interface instead of a "standard" one. So I guess there is still a room for improvement here.

Changing the priority to minor, as the chosen network interface can easily be overriden by configuration.