refactor: split out UDP association handling from serving

[sbruens]

This allows us to re-use the association handling logic in the Caddy server, where the NAT is handled by Caddy.

The NAT handling for the legacy server is split out into its own PacketServe function (to handle PacketConn types) that is similar to our StreamServe function (to handle StreamConn types).

[fortuna]

I believe this needs a redesign. We can make the shared abstraction be an association handler, not a packet handler. It significantly simplifies things because the handler has the association context, rather than having the context of an individual packet only.

And it will likely use less memory, since the handler can call read directly to the buffer it uses.

We will need to extract a NAT component for outline-ss-server.
Note that the handler is still the one to dictate the lifetime of an association, by closing the writes. You can't do that with a packet handler.

[sbruens]

Made the changes to move from a PacketHandler to an AssociationHandler as we discussed earlier. PTAL.

[fortuna]

Thanks for the change. The code is definitely more reusable and composable now!

We should keep an eye on memory consumption. It's not super clear to me if this will have a significant impact on memory usage and allocations. Any ideas on how to ensure it's doing fine?

[sbruens]

We should keep an eye on memory consumption. It's not super clear to me if this will have a significant impact on memory usage and allocations. Any ideas on how to ensure it's doing fine?

The benchmark analysis of the PR looks good. However, I too want to get more confident it performs well under real-world conditions. I'd like to see it deployed and tested in a production-like environment with significant traffic with one of our providers to see what it looks like.