But, in IGV, I can see reads randomly at the transcript
I encountered the same issue some time ago, and it's actually normal to have reads all along the transcripts for 10x single cell data. If you check this question in 10x support forum, you will get that "the fragmentation produces molecules of varying sizes", thus leading to different coverage along the trancripts. So you don't have data that is "3' only", but "3' biased". Try geneBody_coverage.py from RSeQC to have a better idea of the 3' biais on the genome-scale. You can also refer to this preprint - check figure 2 - to go further.