Status update: @ckousik will start on this feature now, as most of the work on js-libp2p-webrtc browser-to-browser is complete (minus the potential multiaddr name change)
@ckousik @ddimaria please use this ticket to capture the work-plan/todos

@p-shahi working branch for this work: https://github.com/ckousik/go-libp2p/tree/ckousik/webrtc-m3

What's gonna be our strategy for finding STUN servers ?
Hard-code some well known servers ? (bad, not local first)

Given STUN is cheap and can be easily rate limited I would argue public go-libp2p servers should include a STUN server, similar to how autonat and relay v2 is implemented right now, then in Kubo I can do the same thing I do for relay v2 where I select nodes based on the DHT keyspace.
I remember discussing this with @mxinden and there were concern than IPFS would unwillingly become public STUN service for all kind of unrelated spammy usecases.

Hard-code some well known servers ? (bad, not local first)

I don't have a good solution here. I would start with this, i.e. I would start with a hard-coded set of servers. Though products would need to make it explicit to the user that they are using the services of company X.