netfilter: conntrack: fix race between confirmation and flush

Commit 5195c14c8b27c ("netfilter: conntrack: fix race in
__nf_conntrack_confirm against get_next_corpse") aimed to resolve the
race condition between the confirmation (packet path) and the flush
command (from control plane). However, it introduced a crash when
several packets race to add a new conntrack, which seems easier to
reproduce when nf_queue is in place.

Fix this race, in __nf_conntrack_confirm(), by removing the CT
from unconfirmed list before checking the DYING bit. In case
race occured, re-add the CT to the dying list

This patch also changes the verdict from NF_ACCEPT to NF_DROP when
we lose race. Basically, the confirmation happens for the first packet
that we see in a flow. If you just invoked conntrack -F once (which
should be the common case), then this is likely to be the first packet
of the flow (unless you already called flush anytime soon in the past).
This should be hard to trigger, but better drop this packet, otherwise
we leave things in inconsistent state since the destination will likely
reply to this packet, but it will find no conntrack, unless the origin
retransmits.

The change of the verdict has been discussed in:
https://www.marc.info/?l=linux-netdev&m=141588039530056&w=2

Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>