Made some RDD dependencies transient to reduce the amount of data needed

to be serialized in closure serialization. This can significantly reduce
the task setup time in Shark when the query involves a large number of
(Hive) partitions.