$S(f) \leq 2$ for all nonconstant $f$.
By context, I understand that you mean factorization in $\mathbb{Q}[x,y]$. If you factor in $\mathbb{C}[x,y]$ then, of course, $x^n-y^n$ is a product of lines. (Daniel McLaury, in comments, mentions a third option, that $y$ might be a fixed integer. I don't know about this.)
Assume $f$ is not constant. Then clearly $x$ and $y$ do not divide $f(x)-f(y)$. Any other linear function can be written as $y-\lambda(x)$ for some affine linear transformation $\lambda(x) = mx+b$ with $m \neq 0$. If $y-\lambda(x)$ divides $f(y)-f(x)$ then $f(\lambda(x)) = f(x)$ for all $x$. Thus, $f(x) = f(\lambda(x)) = f(\lambda(\lambda(x)) = \cdots$. If $\lambda$ is not of either the form $\lambda(x) = x$ or $\lambda(x) = c-x$, then iteration of $\lambda$ has infinite orbits, so this implies $f$ is constant, a contradiction. (If we were factoring over $\mathbb{C}[x,y]$, we'd have to consider $m$ a root of unity other than $\pm 1$.)
Moreover, if $y-(c_1-x)$ and $y-(c_2-x)$ both divide $f(y)-f(x)$, then $f(x) = f(c_1-x) = f(c_1-c_2+x) = \cdots = f(k(c_1-c_2)+x)$ for all $k$, so we can have $y-(c-x) | f(y)-f(x)$ for at most one $c$.
Thus, the most that can divide $f(y)-f(x)$ is $y-x$ and $x+y-c$ for one value of $c$, and $S(f) \leq 2$.
You might have intended to count factors with multiplicity, but this doesn't matter. Let the leading term of $f(x)$ be $f_n x^n$. Then the degree $n$ part of $f(y) - f(x)$, namely $f_n (x^n-y^n)$ is divisible by $x-y$ and $x+y$ only once, so $x-y$ and $x+y-c$ can each divide $f(x)-f(y)$ only once.