ndp: consider exposing Conn constructor which accepts arbitrary PacketConn

[mdlayher]

@stapelberg and I had spoken at GopherCon and I believe he mentioned the need to specify an arbitrary net.PacketConn in this package.

The problem right now is that we effectively force an *ipv6.PacketConn because of multicast groups and control messages. Is this a problem for you, Michael?

[stapelberg]

I can’t see a way to easily test an *ipv6.PacketConn, which is a struct type, not an interface.

Would it be possible to define an interface which includes the methods you need from the PacketConn, so that tests can use a different implementation (to inject/capture packets)?

[mdlayher]

Certainly. Now that I'm thinking about it more, I am also curious if there is real value in giving the caller access to IPv6 control messages as well. I could likely wrap that internally and remove them from the interface as well.