Initializing a ICE session will need to gather candidates. The TURN
can be long to retrieve, and fails can be really long to detect.
There is at least 3 cases of failures:

+ IPv6 badly configured, which can cause a DNS resolution timeout
of several minutes (that's why the IP was cached)
+ Empty DNS entries, causing a resolution failure
+ A TURN server un-reachable or wrongly configured (e.g. 1.1.1.1)

The idea here is to resolve the TURN and test the connection before
caching it. And use it when cached.
This avoid all resolutions steps and we're basically sure that
it was working.

Other approaches:
+ Add a new callback in pjsip to detect that the TURN is taking too long
to remove it for next calls, but I prefer to not add another patch in
pj and it's not an ideal solution
+ trickle ICE to not wait for all candidates, but this is a big changes
and will generate more DHT messages
+ Do not retransmit messages, but this is against the RFC

Change-Id: I2bbc8d9ae76a9e3124c71343df02e2ed077b938d
GitLab: #781