Refactor remotes and handshaking to give every address a fair shot

[nbrownus]

This change unifies the idea of the lighthouse cache and the hostinfo remotes address list. The motivation is:

• To enable a quicker handshake by attempting to transmit the handshake packet to all known remotes on each attempt
• Give every lighthouse an equal opportunity to provide answers, in the event of a buggy or compromised lighthouse, other lighthouses should have their shot at getting the right answer to the client.

Limits are 10 ipv4 and 10 ipv6 per lighthouse (clients) or per client (lighthouse).

Learned addresses changed to only store 1 of each v4 and v6 address. This should benefit hosts that get new public facing ports through a NAT frequently.

This also adds a check to ensure the responder is who we expected at stage 2. There is an open question on whether we keep the result of the work there (less disruptive in the event a lie disrupted a tunnel to a host you were communicating with) or discard the work. There is another PR nearly ready that speaks to this specific problem in a more elegant way.

Given the scope of the change, a few other items were included:

• Dont hold locks in ssh commands #244 is folded in since hostinfo marshaling needed rework.
• Lighthouses will not attempt to roam clients, this keeps a lighthouse on the ip proto the client was configured with.
• Listen address calculation is fixed on linux, sometimes nil:4242 would be reported

[CLAassistant]

All committers have signed the CLA.

[wadey]

All of my issues have been resolved, this is a big change but I think it looks good (and probably safer than what we have now) 👍