Which zUMIs output to use for differential gene expression?

Question

[firestar]

I have single cell Smart-Seq 3 data. The zUMIs pipeline produces about 19 different output count matrices:

umicount
  exon
    all
    downsampled
  inex
    all
    downsampled
  intron
    all
    downsampled
readcount
  exon
    all
    downsampled
  inex
    all
    downsampled
  intron
    all
    downsampled
readcount_internal
  exon
    all
    downsampled
  inex
    all
    downsampled
  intron
    all
    downsampled
rpkm
  exon
    all

Read counts from Exons are the traditional RNASeq counts that one would expect. UMI Counts from exons are the equivalent corrected for PCR duplication. So this will have lower counts compared to reads. The zUMIs paper claims that intron+exon counts improves clustering.

i am interested in views from the experts. What are downsampled counts? What is readcount_internal? What are the implications of using intron+exon? Should rpkm be used at all? Which dataset should be used for differential gene expression? What are the pros and cons of different types of data? What considerations should one keep in mind?