Add (failing) http.rb wrap test for TCPSocket and SSLSocket

[ClearlyClaire]

Test to track #12.

One test fails for the wrap/new issue described there, and another fails for a reason I hadn't stumbled upon so far:

Failure/Error: raise RuntimeError, "Trying to close selector with active monitors: #{@monitors.values.inspect}!"

[coveralls]

Coverage increased (+0.5%) to 86.733% when pulling f37e7ca on ThibG:httprb into 52e165e on socketry:master.

[ioquatix]

The other failure is probably due to http.rb leaking sockets.

[ioquatix]

That looks good - is that the documented to do it? i.e. response.connection.close

How does http.rb work internally, does it have a cache of connections or something?

[ClearlyClaire]

Oops, there are some namespaces issues I didn't notice with the other tests.
The flush things actually fails, it blocks and I'm unsure why… response.connection.close works but it's not something that is explicitly encouraged afaict.

[ioquatix]

Looks like we need to find out how to support

          NoMethodError:
            undefined method `hostname=' for #<Async::IO::TCPSocket:0x0000000002bcc558>

Not sure if this is part of the stdlib for Ruby or not.

[ClearlyClaire]

@ioquatix we kind of support this, it's just that HTTP.rb will call ssl_socket_class.new when your wrapping code expects new to be called internally, and wrap to be called instead!

EDIT: To be more precise, Async::IO::SSLSocket isn't a drop-in replacement, as creating it with new as one would usually do will result in it getting a socket instead of the wrapper it expects. Therefore, the subsequent call to hostname= will be on the socket and not on the wrapper/OpenSSL's SSLSocket.

[ioquatix]

I guess we need to consider how this should work. Whether we should provide some specific Async::IO::Wrap::SSLSocket or implement a constructor for Async::IO::SSLSocket which multiplexes in the same way as TCPSocket.

Because performance of SSLSocket is important, I hesitate to make the constructor complicated for this edge case. But maybe a special wrapper might be appropriate. Both options are kind of annoying.

[ioquatix]

I am going to see if I can make this work in a way I'm happy with. Thanks for contributing a failing spec.

[ioquatix]

I've been playing around with this and while it's pretty straight forward with TCPSocket.open, it's less trivial since SSLSocket doesn't support the same method. I feel like it would be nice if they both had a similar open method which would simplify the internal API quite a bit.