limma never assumes that any factor or variable has the same variance between genes. In other words, everything is assumed to be heteroscedastic. Normally we don't think about variances of factors anyway, only about that of expression values.
I am somewhat puzzled why you would ask this question and why you think that something special needs to be done to model heteroscedasticity. I assume there is something more to your question that you haven't told us.
All the RNA-seq methods -- edgeR, DESeq2 and voom -- allow the variance to depend on the mean for each gene. There is no substantial differences between the methods in that respect. voom is slightly more flexible in the terms of the mean-variance relationship than the others (when the library sizes are inconsistent), as discussed in the Discussion section of the voom paper in Genome Biology.
If you want to allow variances to depend on an experimental factor, this can be done using voomWithQualityWeights(). See the article cited in the help page for details.