Remove temporary channels

[pwojcikdev]

Our current networking design is strange in this aspect, that we treat connections accepted by our node (temporary channels) differently to connections that are initiated by our node. There is no fundamental reason for such behavior, TCP allows for bidirectional communication and doing so may lead to more robust network, since some peers aren't directly reachable (NAT, firewalls).

The logic for handling those temporary channels was also not great, it created a new channel for each processed message which led to inefficiencies, especially when paired with fair queuing. Look at nano::transport::tcp_channels::process_message (...). It also used node id for channel lookup when processing messages. Here I'm associating each realtime connection with a single channel, without relying on node id, which is simpler and more robust approach.

This PR moves management of lower-level connections fully to tcp_listener and implements them using coroutines. Each connection attempt is given a timeout (by default 60 seconds) and node id conflicts are only rejected for same IP addresses.

Previous attempt to fix this issue was #3928

[clemahieu]

How does changing the order avoid ephemeral ports?

[pwojcikdev]

The way tests setup nodes is that second node connects to the first one. Connection is initiated from ephemeral port to a specific listening port. Here we can only predict the listening port, so node needs to connect to node2.

[pwojcikdev]

If there is a better way I'll be happy to change it. I wanted to get it working without removing asserts.