Even for UDP sockets, we should handle EAGAIN errors because it can happen.. Apparently recvfrom can return EAGAIN on udp sockets if for example there's a bad checksum in the packet or some other similar error

Fix a crash in case a relay candidate could not be added (because of an error or because it's a redundant candidate), do not try to refresh it

Once we process the pending checks, we should free them to avoid reprocessing them again if a dribble-mode set_remote_candidates is called

Also, do not re-create it if it already exists

use sizeof struct sockaddr instead of sizeof sockadddr_storage for doing the bind.. this fixes stund on BSD systems. Thanks to Tom Murphy for the fix

Do not use connect on UDP sockets and use sendto/recvfrom instead of send/recv because apparently on BSD a connect will fail on UDP sockets

Only keep the srv-reflexive candidates alive by resending our initial STUN discovery during keepalive phase... this should fix the sending of over 300 completely useless connchecks that just spammed the network and filled our max saved ids for the stun agent requests, thus rendering future connchecks useless