Quinn endpoints leak causes increased memory on public peers and eventual refusion to accept new peers

[LeeSmet]

It seems that sometimes an Endpoint (possibly wrapped in some type) can leak. This increases memory consumption on the node, and eventually, if enough of them leak (65K), no new quic connnections can be created.

[LeeSmet]

After more debugging, it turns out the Incoming mentioned in the errors are actually still stored in the internal quinn queue for new connections. The problem is that since accepting a connection then waits for client data, the clients can stall the server by not sending the handshake. Potentially this could be triggered by random udp packets as well (not sure).

To solve this, we need to wrap the quic setup (after the initial accept) in a timeout, and potentially also resolve multiple peers side by side to avoid a client not finishing the handshake and delaying others.