Make get_request timeouts recoverable

[Juerd]

With short select() delays (e.g. in event loops) and slow connections,
the client may not have more data ready before select timeouts.

This change lets get_request return a different kind of false (defined
instead of undef) in case of a select timeout, so that the caller can
choose to call it again. To make this work, the same buffer
is re-used with _need_more every time.

Hacks that inherit from HTTP::Daemon and override _need_more, will
break, as will existing code that explicitly tests the result of
get_request for definedness rather than its boolean value, and then
assumes that it will be an HTTP::Request object. Both cases are
unlikely.

The unit test is based on chunked.t with a lot removed.

[midenok]

Why you need two select()'s: first one in your event loop and second one in _need_more()? select() in _need_more() is the implementation of timeout logic which is not needed in event loop (as far as you've added HTTP::Daemon::ClientConn fd to it). Instead of vague return type and small timeout hack you should avoid select() in _need_more() completely and check for EAGAIN after sysread(). There should be some option to tell HTTP::Daemon to don't do select(), e.g.:

  $d = HTTP::Daemon->new(
           NonBlock => 1
       );