This change modifies the Yarn module so that all the logic related
to running the ApplicationMaster is localized. Instead of, previously,
4 different classes with mostly identical code, now we have:

- A single, shared ApplicationMaster class, which can operate both in
  client and cluster mode, and substitutes the old ApplicationMaster
  (for cluster mode) and ExecutorLauncher (for client mode).

The benefit here is that all different execution modes for all supported
yarn versions use the same shared code for monitoring executor allocation,
setting up configuration, and monitoring the process's lifecycle.

- A new YarnRMClient interface, which defines basic RM functionality needed
  by the ApplicationMaster. This interface has concrete implementations for
  each supported Yarn version.

- A new YarnAllocator interface, which just abstracts the existing interface
  of the YarnAllocationHandler class. This is to avoid having to touch the
  allocator code too much in this change, although it might benefit from a
  similar effort in the future.

The end result is much easier to understand code, with much less duplication,
making it much easier to fix bugs, add features, and test everything knowing
that all supported versions will behave the same.

Made some tweaks to the YarnAllocator interface to cover both
APIs more easily. There's still a lot of cleanup possible on
that front, but I'll leave that as a separate task.

This avoids the NPEs that would happen if code just kept going.

Makes code a little cleaner and easier to follow.

Conflicts:
	yarn/alpha/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala
	yarn/alpha/src/main/scala/org/apache/spark/deploy/yarn/ExecutorLauncher.scala
	yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala
	yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/ExecutorLauncher.scala

Avoid an interrupted exception when shutting down from the
reporter thread. Also, make extra effort to finish the RM
connection.

Conflicts:
	yarn/alpha/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala
	yarn/alpha/src/main/scala/org/apache/spark/deploy/yarn/ExecutorLauncher.scala
	yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala
	yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/ExecutorLauncher.scala

This change also avoids overriding the app's status with "SUCCEEDED"
in cluster mode when the shutdown hook runs, by signaling the AM that
the SparkContext was shut down (see YarnClusterScheduler.scala). That
way the AM can correctly expose its final status to the RM.

Conflicts:
	yarn/alpha/src/main/scala/org/apache/spark/deploy/yarn/ExecutorLauncher.scala
	yarn/alpha/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala
	yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/ExecutorLauncher.scala