This is cool, but missing a LOT of details between steps 4 and 5, which is the meat of the quicksort. Actually, the first and last elements of step 4 would be swapped, which means the order depicted in step 5 is incorrect.
I'd guess if you care more about speed than memory it might be faster to just move elements into new array - sequence through old array appending to start/end of new array according to pivot comparison. You'd be moving every element vs leaving some in place with a swap approach, but the simplicity of the code & branch prediction might win out.
shame, shame, I should have double-checked before posting.