In most cases it seems protocols initially described in the context of composite order groups can be converted to prime order groups, thanks to the work of David Freeman and others. Which is just as well as calculating pairings over composite order groups is impossibly slow, in part because only a few pairing-friendly curves of low embedding degree are suitable for use with composite order groups. So the answer to the second part of your question is that (hopefully) composite order groups for bilinear maps are never needed (although they might be continue to be used in academic papers to initially describe a new protocol).

In most cases it seems protocols initially described in the context of composite order groups can be converted to prime order groups, thanks to the work of David Freeman and others. 

Which is just as well since calculating pairings over composite order groups is impossibly slow, in part because only a few pairing-friendly curves of low embedding degree are suitable for use with composite order groups. So the answer to the second part of your question is that (hopefully) composite order groups for bilinear maps are never needed (although they might be continue to be used in academic papers to initially describe a new protocol).