Use filesystem to collect RDDs in PySpark.

Passing large volumes of data through Py4J seems
to be slow.  It appears to be faster to write the
data to the local filesystem and read it back from
Python.