The previous version assumed that a CLASSPATH environment variable was
set by the "run" script when launching the process that starts the
ExecutorRunner, but unfortunately this is not true in tests. Instead, we
factor the classpath calculation into an extenral script and call that.

NOTE: This includes a Windows version but hasn't yet been tested there.

1) Add support for HADOOP_CONF_DIR (and/or YARN_CONF_DIR - use either) : which is used to specify the client side configuration directory : which needs to be part of the CLASSPATH.

2) Move from var+=".." to var="$var.." : the former does not work on older bash shells unfortunately.

See the JIRA for more details.

I was only able to test the bash version (don't have Windows)
so maybe check over that the syntax is correct there.

- we do not need getEnvOrEmpty
- Instead of saving SPARK_NONDAEMON_JAVA_OPTS, it would be better to modify the scripts to use a different variable name for the JAVA_OPTS they do eventually use

other issues

Conflicts:
	run2.cmd

Changed teh scala version to 2.9.2, so that the classes can be found when the classpath is expanded.

- Bundle Py4J binaries, since it's hard to install
- Uses Spark's `run` script to launch the Py4J
  gateway, inheriting the settings in spark-env.sh

With these changes, (hopefully) nothing more than
running `sbt/sbt package` will be necessary to run
PySpark.