Yes, it's possible to perform an in-place merge sort. That does not mean it doesn't require any extra memory, since the recursion still takes O(lg n) stack space (just like quicksort).
Edit: it's been long since I read the Katajainen paper, and I never paid attention to the space complexity claim; apparently, merge sort can be done in O(1) extra space.
But no, all this does not mean one algorithm is better than another, unless you specify what you mean by that word. Randomized quicksort, while it can go quadratic, still performs a smaller expected number of operations, regardless of the input. Which one "works better" for your data depends on factors ranging from the statistics of the data to the hardware the program runs on.