If sending a lot of data and/or the other peer is not reading it, then
socket buffers can overflow. This is communicated from the kernel by
returning EAGAIN. In GIO, it is modelled by g_output_stream_write()
and g_socket_send_message() returning G_IO_ERROR_WOULD_BLOCK.

It is also problematic that that we're using synchronous IO in the
shared GDBus IO thread. It means that one GDBusConnection can lock up
others.

It turns out that by porting from g_output_stream_write() to
g_output_stream_write_async() we fix the EAGAIN issue. For GSocket, we
still need to handle things manually (by creating a GSource) as
g_socket_send_message() is used.

We check the new behavior in Michael's producer/consumer test case (at
/gdbus/overflow in gdbus-peer.c) added in the last commit.

Also add a test case that sends and receives a 20 MiB message.

Also add a new `transport' G_DBUS_DEBUG option so it is easy to
inspect partial writes:

 $ G_DBUS_DEBUG=transport ...