errno doesn't contain anything useful in this case

otherwise we may end up in a weird state where we're counting connections that we don't have or have cleared a timeout without a reason.

it should call conn_free to cleanup the connection struct now.

so that if it fails the user is promptly notified. This requires us to make sure that fds 0-2 are opened, so steal the same trick slowcgi have.

instead, drop the current connection (we can't do anything about it) and try to continue to work. OOM can be a temporary situation, and in every case it's not nice to terminate other in-flight connections. To do so, refactor the freeing of the connection into conn_free, which now also looks for the state of the timeout. In every case we reach conn_free the timeout is not pending, but it doesn't hurt to check for it anyway. Future changes may want to free early.