You're misunderstanding. They don't explain how to do partitioning at all. Step 3 is just tagging the elements that are above the partition element, which there happen to be two of.
Well, to the extent a picture can explain they say choose at random. That's what the dice shown is about.
A random pivot is... fine. It can't solve the worst case perf problem, and it won't ensure high performance for other cases either, but hey you did pick a pivot and this algorithm is often fast enough.