The sources I have seen use a definition consistent with a single variable. As for the wolfram link, I think the wolfram definition is suggestive of defining the term for a single variable since it refers to inverses and repeated compositions (though this phrasing technically allows for the possibility of self maps on $\mathbb{C}^n$).
The question then becomes: Why can't we call more functions "elementary"? This question is explored here. Of course, nothing is stopping us from calling other functions, such as polynomials in several variables, "elementary." The question is: How useful is it to do so?
For instance, our defined notion of "elementary functions" enjoys application in differential algebra. These functions are closed under differentiation, but not under antidifferentiation. Liouville's theorem gives an answer to which kind of elementary functions have elementary antiderivatives, and the Risch algorithm makes this theorem practicable.