Recently I've been wondering if the numreads from samtools coverage could be used as count tables for posterior differential expression analysis (DESEQ2, EdgeR), the results would be different from using counts from FeatureCounts?

If you are going to go through the trouble of defining regions corresponding to each exon and then adding those counts up (when using samtools coverage), why not stick with featureCounts that understands splicing and will make use of the gene models in the GTF file. samtools coverage would also not know about multi-mapped reads which need to be properly addressed during counting.

the results would be different from using counts from FeatureCounts?

Yes results would be different and perhaps not logical (unless you are dealing with non-overlapping non-spliced genes with no multi-mapped reads).

Use correct tool meant for a specific analysis. Because samtools overage is going to give you counts they are not necessarily going to be usable as is for DE analysis.