This should go into 1.0 since it would return wrong data when the bug happens (which is pretty likely if cancellation is used). Test case attached.

1. Do not put partially executed partitions into cache (in task killing).

2. Iterator returned by CacheManager#getOrCompute was not an InterruptibleIterator, and was thus leading to uninterruptible jobs.

Thanks @aarondav and @ahirreddy for reporting and helping debug.

Author: Reynold Xin <rxin@apache.org>

Closes #521 from rxin/kill and squashes the following commits:

401033f [Reynold Xin] Merge branch 'master' of https://git-wip-us.apache.org/repos/asf/spark into kill
7a7bdd2 [Reynold Xin] Add a new line in the end of JobCancellationSuite.scala.
35cd9f7 [Reynold Xin] Fixed a bug that partially executed partitions can be put into cache (in task killing).