Fix TCP error handling in conn_poll

[Sean-Der]

Don't kill the entire connection because TCP has errored

[paullouisageneau]

I think it works but it is not efficient since instead of marking the pair as failed immediately if the TCP connection fails, it waits for timeout anyway.

You should at least fix the number of retransmissions for TCP. currently it seems to me that agent_arm_transmission() sets it to 6 like for UDP, meaning you wait for a long time and call conn_tcp_connect() multiple times uselessly before timing out.

[paullouisageneau]

I'm merging this but it looks like there are still issues introduced by TCP support, for instance a partial message makes the polling thread basically busy-wait and use 100% CPU, and it crashes when TCP connects if an ICE server is set (fixed by #316).

[Sean-Der]

Thank you @paullouisageneau

Do you have recommendations/ideas on unit tests or anything else I can do to catch more stuff?

[paullouisageneau]

@Sean-Der I looked into it and ended up making a bunch of fixes in #317. The most important one is that the ICE-TCP framing protocol must be implemented with buffering, since as soon as the TCP buffer is full send() may result in a short write. I'm not finished, as there are still some issues like retransmissions happening over TCP, I'll be back to it soon.