CoalescedRDD for many partitions

Remove try/catch block that can't be hit.

call setConf from input format if it is Configurable

CoalescedRDD changed to work with a big number of partitions both in the original and the new coalesced RDD.
The limitation was in the range that Scala.Int can represent.

Add PairRDDFunctions.subtractByKey.

Fix local metrics

Add a log4j compile dependency to fix build in IntelliJ

Also rename parent project to spark-parent (otherwise it shows up as
"parent" in IntelliJ, which is very confusing).

Close the reader in HadoopRDD as soon as iteration end.

Fix broken link to YARN documentation

Fix broken link in Quick Start

Send block sizes as longs in BlockManager updates

This also refactors out the common "gotNext" iterator pattern into
a shared utility class.

Adds choice of persistence level to Bagel.

Conflicts:
	core/src/test/scala/spark/scheduler/DAGSchedulerSuite.scala

Fix reference bug in Kryo serializer, add test, update version

Notify standalone deploy client of application death.

Conflicts:
	core/src/main/scala/spark/storage/BlockManagerMaster.scala
	core/src/test/scala/spark/scheduler/DAGSchedulerSuite.scala

Usually, this isn't necessary since the application will be removed
as a result of the deploy client disconnecting, but occassionally, the
standalone deploy master removes an application otherwise.

Also mark applications as FAILED instead of FINISHED when they are
killed as a result of their executors failing too many times.

Use the LocalSparkContext abstraction to properly manage clearing
spark.driver.port.