Refer to the JIRA for the design doc and some perf results.

I wanted to call out some of the more possibly controversial changes up front:
* Map outputs are only stored in serialized form when Kryo is in use.  I'm still unsure whether Java-serialized objects can be relocated.  At the very least, Java serialization writes out a stream header which causes problems with the current approach, so I decided to leave investigating this to future work.
* The shuffle now explicitly operates on key-value pairs instead of any object.  Data is written to shuffle files in alternating keys and values instead of key-value tuples.  `BlockObjectWriter.write` now accepts a key argument and a value argument instead of any object.
* The map output buffer can hold a max of Integer.MAX_VALUE bytes.  Though this wouldn't be terribly difficult to change.
* When spilling occurs, the objects that still in memory at merge time end up serialized and deserialized an extra time.

Author: Sandy Ryza <sandy@cloudera.com>

Closes #4450 from sryza/sandy-spark-4550 and squashes the following commits:

8c70dd9 [Sandy Ryza] Fix serialization
9c16fe6 [Sandy Ryza] Fix a couple tests and move getAutoReset to KryoSerializerInstance
6c54e06 [Sandy Ryza] Fix scalastyle
d8462d8 [Sandy Ryza] SPARK-4550