Add Eio_unix.Net module

[talex5]

• Add Eio_unix.Net module and move network-related items from Eio_unix to there.
• Replace Eio_unix.socketpair with Eio_unix.Net.socketpair_{stream,datagram}. Previously, it took a type argument (stream or datagram), but always returned a stream_socket.
• Move sockaddr conversions to eio.unix so they can be shared.
• Implement getnameinfo in eio.unix to share it by default.
• Eio_linux.Low_level.send_msg no longer throws away the number of bytes sent.
• Add close to all socket types.
• The destination in send is now optional, for unconnected datagram sockets.
• send now takes a iovec, not just a single buffer. The stream API allows this and we tested it for datagrams using that API due to the previous type confusion.
• Add Unix_datagram socket address type for completeness.
• Add send_msg and recv_msg stubs to Eio_posix. This avoids converting to bytes, and will be useful to add FD passing later.

I was planning to include FD passing in this FD, but it got too long so I've split that out for later.

Fixes #342.

[talex5]

Having tried this out for a bit, this API is a little annoying because we also need unix equivalents to Net.run_server, etc. So I think it might be better to make the main Eio types polymorphic instead.

[talex5]

I've removed the FD passing bits for now and put them at https://github.com/talex5/eio/tree/fd-passing, which I'll submit separately once this is merged. So this PR is just adding a skeleton Eio_unix.Net and sorting out the stream/datagram confusion now, which should make it easier to review.

[talex5]

I've pushed a couple of minor changes:

• I reused the Unix tag for datagrams, since it's the same file either way, and therefore the same address. I think Windows named pipes should have their own tag if we want them, but that should be a separate issue.
• I added Eio_unix.Net.import_socket_datagram to match import_socket_stream.

If there are no objections I'll merge this tomorrow, so I can then submit the FD-passing PR.

[avsm]

One implication of this change is that Eio.Net.accept_fork and other connection handlers now must implement close (whereas previously it was just a Flow.two_way).

[avsm]

(For reference, this is the diff I applied to bring other libraries uptodate with this change https://github.com/avsm/eeww/commit/dfdd8f248b249254b63bc5f18f8ab172ec94ca0e)

[talex5]

Good point. At the moment, accept_fork always closes the flow when the handler finishes, so if you close it manually then that will fail, which is why we didn't have close here before. Perhaps we should revert the addition of close here.

[avsm]

It would be cleaner to just pass a Flow.two_way I think, and leave closing to the connection handler.