Handle 408-response connection closes. #146

PR #145 ensures that we're checking for disconnections when we reacquire connections, by using the StreamReader.at_eof interface.

However @njsmith points out, that's not always quite sufficient. #143 (comment)

There's StreamReader.at_eof, but that will fail if the HTTP server is polite and sends a 408 Request Timeout before it closes the connection. I think the best you can do is:

def asyncio_stream_is_readable(stream):
     return stream._buffer or stream._eof

or else giving up on the Stream API and using protocols/transports directly.

Either of those would be do-able.

How common is 408 usage in the wild - are there example URLs we can confirm against?

Also worth noting that there's also always going to be a race network condition where we start getting a 408 response that was about to be sent, just as we're in the process of sending out a new request.