Eio port

[talex5]

This switches capnp-rpc from Lwt to Eio. One particularly nice side effect of this is that Service.return_lwt has gone, as there is no distinction now between concurrent and non-concurrent service methods.

Notes:

• Some of the "unix" functions can be moved into the core library with Eio.
• Mirage support is gone. The main library should now work with both Unix and unikernels.

This commit aims to keep the diff (relatively) small. Several more things are needed after this is merged. In particular:

• Performance improvements (such as sending multiple messages at once).
• Handling leaked refs correctly. We need to schedule a callback on our owning domain's event loop from the finalizer. Eio doesn't currently provide a way to do this.

[tmcgilchrist]

Can we keep the LWT version available for OCaml 4.14? We will need to move services to OCaml 5 and EIO over a longer period of time, both for platform support reasons and just because it will take some time to port things over.

[talex5]

Can we keep the LWT version available for OCaml 4.14?

The existing release will continue to work there. capnp-rpc is pretty stable, so I'm not expecting any major new features that would need back-porting.

[tmcgilchrist]

We'll take a release of this when it's ready. clarke and solver-service could use it.

[patricoferris]

We might need to wrap this to catch

Unix.Unix_error(Unix.EOPNOTSUPP, "setsockopt", "")

Stumbled across this new change when using this branch on macOS with a unix domain socket.

[talex5]

Sounds good. Not sure whether to do it here, or get ocaml-multicore/eio#575 finished. We could then do something like:

Eio.Net.connect ~sw net eio_addr ~options:[Try (Delay false)]

[RyanGibb]

Wrapping this in a try stops us crashing when a flow has already been closed. I.e.

  try
    Eio.Flow.shutdown t.flow `All
  with | Eio.Io (Eio.Net.E Connection_reset _, _) as ex ->
    Log.info (fun f -> f "%a" Eio.Exn.pp ex);

[talex5]

Everything here got merged with #284 and #287, except for disabling Nagle's algorithm, which is probably no longer necessary. At least, it no longer seems to affect the benchmark results.