Spark runner batch translator to work with Datasets instead of RDDs

Currently, the Spark runner translates batch pipelines into RDD code, meaning it doesn't benefit from the optimizations DataFrames (which isn't type-safe) enjoys.

With Datasets, batch pipelines will benefit the optimizations, adding to that that Datasets are type-safe and encoder-based they seem like a much better fit for the Beam model.

Looking ahead, Datasets is a good choice since it's the basis for the future of Spark streaming as well (Structured Streaming) so this will hopefully lay a solid foundation for a native integration between Spark 2.0 and Beam.