Yes, you can use q2-cutadapt to trim your demultiplexed reads prior to denoising. See this tutorial
Yes that is the point — to collapse at species level (since OTUs from different primer regions will be unique even if they overlap). But I think we can all agree this is not the best solution for your case.
Regarding your questions about q2-fragment-insertion, you can check out this post about this community plugin, or perhaps @Stefan can answer your questions.
I hope that helps!