Support the fd socket option when listening

[tomas-abrahamsson]

This PR allows the user to set the {fd,Fd} socket option when starting a grpc server. One use for this is if we want to open on any free port number, but may want to control the ip address. Another use case could be if the fd was opened by an external program with more privileges and sent to erlang using sendmsg(2), for instance with fd_server.

I'll try to describe our context a bit which led us to this proposal. We start grpc servers dynamically and, depending on our user, we may need to start several grpc servers. We may have to stick to some particular ip address, and we are not alone on the server, so we can often not bind to all addresses. But we can use any available port number. So we would ideally like to set the port to 0 to let the operating system assign us some available port. We have some means to communicate the dynamically selected port number to the user.

First we tried to set the ip option and the port => 0 option, but we faced a couple of other issues: one is that the port number is used as part of a process name and we would like to be able to open more than one grpc server for a particular ip address, another issue was that we would need to somehow get hold of the socket to learn what port number was actually assigned by the operating system.

The fd option could solves both of these issues: we can open the listen socket in advance, use inet:sockname/1 and set the ip and port options. Just gen_tcp need to be convinced to not try to bind, and this is what this PR does.

Do you think this could be ok?

[tsloughter]

I still need to look this over fully but this is definitely ok and glad to have it. This is also used by services that use systemd's feature to get passed an open socket as fd 3.

I'd also like to see unix domain socket support added, maybe I'll get that done soon.

[tomas-abrahamsson]

I hope I fixed the test suite issue. (I first had some troubles running the tests locally, but just managed to get them going. It turned out the port used in the test suite was occupied, by a locally running Jenkins of all things, haha :) Once I stopped that, all the weird ssl errors went away nicely)

[tsloughter]

Was about to merge this when I saw the end for testcase unary_garbage_collect_streams was removed, I take it that was on accident? :) Can you add it back and I'll merge.

[tomas-abrahamsson]

Sorry, yes, absolutely accidental. Good catch, thanks. Fixed.

[tomas-abrahamsson]

I should mention that I've noticed occasionally error logger printous about stealing control of fd= from output driver tcp_inet #Port, or something similar. I don't have the exact printout at hand at the moment, and fail to reproduce it locally, so this is from memory. There seems to be no ill effects, except for the printout itself, but maybe I ought to chase this a bit to see if it can be avoided?