pipeline interface

queue operations paired with a future (`reply Lwt.u` or `reply -> unit` possibly, depending on backend); send the whole batch in one go. Possibly with the following API:

```ocaml
Redis.pipelined client
  (fun () ->
     R.set c "a" 1;
     R.set c "b" 2)
```

where normal operations could take an optional `reply IO.future` that they fullfill once the pipeline is processed.