Inequality for Distribution of points in space

No, this is not true, here is the sketch of a sequence of distributions where your sum asymptotically grows slower than $n^3$. In the following I am using the symbol $\asymp$ for expressions whose ratio is bounded above and below by some positive constants independent of $n$, and similarly $\lesssim$ for inequalities which hold up to a positive constant.

Start with the following construction: Let $R>1$ be an integer, and let $x_{k,l} = (k,l)$ be the points with integer coordinates $k,l \in \{-R,\ldots,R\}$. These are $n=(2R+1)^2 \asymp R^2$ points with $$ \sum_{k,l=-R}^R \|x_{k,l}\|^2 = \sum_{k,l=-R}^R (k^2+l^2) \asymp R^4 \asymp n^2. $$ Furthermore, for any fixed $k_0,l_0 \in \{-R,\ldots,R\}$ we have $$ \sum_{{k,l=-R}\atop{(k,l) \ne (k_0,l_0)}}^R \frac{1}{\|x_{k,l} - x_{k_0,l_0}\|^2} \le \sum_{{k,l=-2R}\atop{(k,l) \ne (0,0)}}^{2R} \frac{1}{\|x_{k,l}\|^2} \le \sum_{m=1}^{2R} \,\, \sum_{\max(|k|,|l|)=m} \frac{1}{\|x_{k,l}\|^2}, $$ which is rearranging the sum into sums over boundaries of concentric squares of radius $m$. The number of integer points on such a boundary (i.e., the number of terms of the inner sum, given the radius $m$) is $\asymp m$, and each term in that sum is $\frac{1}{\| x_{k,l} \|^2} \asymp \frac{1}{m^2}$, so we get that the total sum above is $$ \asymp \sum_{m=1}^{2R} \frac{1}{m} \asymp \ln R \asymp \ln n. $$ Summing this up over all lattice points $(k_0,l_0)$, we get that $$ \sum_{{k,l,k_0,l_0=-R}\atop{(k,l) \ne (k_0,l_0)}}^R \frac{1}{\|x_{k,l} - x_{k_0,l_0}\|^2} \lesssim n \ln n. $$ For your functions $f_i$, where $i$ now runs through all pairs $(k,l)$ as above, we then get $$ \sum_{k,l=-R}^R f_{k,l}(x_{-R,-R},\ldots,x_{R,R}) \lesssim n^4 + n \ln n, $$ which obviously is not quite a counterexample yet. However, from the homogeneity of the terms it is immediate that scaling by a factor $\lambda > 0$ leads to $$ \sum_{k,l=-R}^R f_{k,l}(\lambda x_{-R,-R},\ldots,\lambda x_{R,R}) \lesssim \lambda^2 n^4 + \lambda^{-2} n \ln n. $$ A choice of $\lambda = n^{-3/4} (\ln n)^{1/4}$ (so that both terms have the same asymptotic behavior) leads to $$ \sum_{k,l=-R}^R f_{k,l}(\lambda x_{-R,-R},\ldots,\lambda x_{R,R}) \lesssim n^{2.5} \sqrt{\ln n}. $$ Somehow I suspect that this is actually the lower bound for the asymptotic behavior of your sum for general $(x_i)$, but I don't have a proof of that.