You're right: for differential expression itself, I don't believe that Cufflinks is ideal due to the fact that it doesn't handle the wide variation in counts that can exist. One can end up with log base 2 fold change difference of upward of 100, which is astronomical and doesn't make much sense. You are not required to use FPKM with Cufflinks, though; If you must use Cufflinks, for whatever reason, then use geometric normalisation.

DESeq2 handles these issues very well and allows you to perform a regularised log transformation (rlog) or variance stabilised transformation (VST) on your data, which produces very nice P values and fold-changes. So, many people then try to extract raw counts from the TopHat/Cufflinks method and then do the differential expression in DESeq2, but this is not possible directly using TopHat/Cufflinks.

A much simpler way to analyse RNA-seq data is with Kallisto ( https://pachterlab.github.io/kallisto/ ), which extracts raw counts over a CDS FASTA that you supply. If you supply GENCODE's CDS FASTA ( https://www.gencodegenes.org/releases/current.html ), then you can get count abundances over upward of 200,000 transcripts and their isoforms.