multiplexed_socket: fix race condition on acceptation

Because some callbacks can take a long time (accepting a SIP channel
will starts to sync profiles, ask for new channel, etc), control packets
are handled in their own thread. However, if the peer accepts a request,
it can starts to use it and send data on it. When finished, it will close
the socket and both peer will remove the channel.
In some case, the channel can be removed before
MultiplexedSocket::Impl::onAccept, causing upper layers to handle a
bad channel. In this patch, we removes the channel only when upper layers
know that the channel is accepted and ready to use. Also, onAccept()
doesn't create wrong channels (which was a bug, where isInitiator() was
in the wrong state).

GitLab: #659
Change-Id: I4f45feacc2624ce0135ace09f7e0188b56fcf81f