Tungsten execution engine

Tungsten stores rows off-heap and generates JVM bytecode for whole stages so Spark SQL is not a naive iterator of JVM objects.