Dynamic super-peers worked for (old) Skype didn't it?
For small webrtc conferences, treating certain peers as an SFU can certainly work. However, bandwidth requirements would be substantially higher than full-mesh for the supernodes, and equivalent to SFU peers for non supernodes - so supernodes would need to be chosen wisely (and you would still be very limited on number of participants - then you are entering the realms of peering supernodes and intelligent routing of streams). Additionally, this still wouldn't be e2e encrypted. Its just a (current) limitation of webrtc.