Right now we strcmp() the isSecure and isHttpOnly values,
but if they are missing we will crash since strcmp() is
not NULL safe. Use g_strcmp0() instead.

The fix introduced in ef7fc058 for bug 578990 caught most cases where
we'd write a request to an already-closed connection, but not all.
Unfortunately, at some point the code path for recovering from that
case got broken, so when it did happen, we'd always end up returning
the error to the app rather than retrying the request. This is now
fixed again.

The specific recent cause of the problem was that the
multiple-pending-connections fix (baa316ac/bug 594768) results in us
sometimes opening more connections than we need, and so some of those
connections become idle without ever having had a message sent across
them. Servers tend to have different timeouts/behaviors here than with
idle persistent connections, so we should time them out more
aggressively. So now we close a connection after 3 seconds if it
hasn't been used yet.

Noted by Enrico Tröger on IRC.

We use GResolver so we don't need to check for getaddrinfo() etc
ourselves, and we don't need to check for IPv6 support, since
GResolver/GSocket require it unconditionally.

Fixes both https://bugzilla.gnome.org/show_bug.cgi?id=587709 (the
current hack for letting you autogen without libgcrypt.m4 does the
wrong thing if you have gnutls-devel but not libgcrypt-devel
installed) and https://bugzilla.gnome.org/show_bug.cgi?id=592275
(older autotools don't automatically create the m4/ dir).

noted by Nguyen Thai Ngoc Duy in bug 581342

Only allowing a single connection attempt at once really slows down
pages with lots of subresources on servers that disallow persistent
connections.

https://bugzilla.gnome.org/show_bug.cgi?id=594768