Fix use of lwt_io + conduit-tls

[dinosaure]

According to the documentation of conduit-tls, the implementation with ocaml-tls is not thread-safe, something like both (send flow) (recv flow) is unsafe. It's mostly about the wrapping of a Conduit.flow with Lwt_io - because the flow can be a TLS flow, we must protect recv and send with a mutex anyway.

The question behind such patch is: should we protect conduit-tls with a mutex - and by this way, we are not mandatory to protect any flow with a mutex or should we keep this design? To this existential question, we should make an issue a talk about this issue later - note that this is not a part of the core of conduit.

The second patch is about the first-read behavior of Lwt_io. I'm not sure how we can properly schedule that but it seems that the Lwt_io wants to read first when we start a TLS connection - I suspect that it schedule the action from the underlying state of the socket and, because we receive a part of the handshake, it considers Conduit.recv as the first operation. However, for the HTTP/1.1 point-of-view, this is not true when the server expect to receive something - so client and server expect to receive something at the end.

conduit-tls is able to produce an opportunity to schedule Conduit.send even if we called Conduit.recv, it returns Input 0 (which means that we received nothing, but the connection still is established). On the other side, recv given to Lwt_io can yield to give the opportunity to the application to scheduler Conduit.send/Lwt_io.write.

Finally:

1. the question about conduit-tls is a bit hard but an issue will be done which describe precisely the situation
2. the problem with lwt_io is an inheritance of 1) and it seems that tls.lwt fix it in its own way

[dinosaure]

Ok, it seems that the first-recv behavior of Lwt_io is true in any case (TLS or not), the second commit gives the opportunity to reschedule Conduit.send with an Input 0.

[dinosaure]

The real conclusion of this patch is: don't use Lwt_io when you need a full control of how to schedule recv/send...