If multicast routing software configures kernel to forward incoming
multicast frames from external interfaces to lo, then, currently, every
packet will be delivered to local host exactly twice:

- once via xmit over real loopback device, and
- once via dev_loopback_xmit() called by ip_mc_output.

-> Fix this double delivery by avoiding second dev_loopback_xmit() call if
target device is already known to be of loopback kind.

For the referece: IPv6 already behaves this way starting from Linux 2.1.89
(commit 3b0db29e in historical repository):

    ip6_finish_output2(net, sk, skb)
    	if (ipv6_addr_is_multicast(daddr)) {
    		if (!(dev->flags & IFF_LOOPBACK) && sk_mc_loop(sk) &&	<-- NOTE `dev->flags & IFF_LOOPBACK`
    			...
    			dev_loopback_xmit

    https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/net/ipv6/ip6_output.c?id=v5.18-rc5-28-ga7391ad35724#n81

XXX link to my test program

XXX not sure whether we should `& IFF_LOOPBACK` or check `dev == net->loop_device`
    or `& IFF_LOOPBACK && dev->netns == net`

XXX test

XXX Cc
Signed-off-by: Kirill Smelkov <kirr@nexedi.com>