Plug VertexProgram iteration leak on empty Spark RDD partitions

If SparkExecutor gets an RDD with empty partitions, it can invoke VertexProgram.workerIterationStart without ever invoking VertexProgram.workerIterationEnd.

For vertex programs that allocate and release meaningful resources in the start/end methods, this can lead to resource leaks.

I already tested a fix that I made against the 3.2 series. I will submit PRs momentarily.