This is expected to fix the old in6_selecthlim() panics. The nature of
the panic is that a packet sending thread will obtain the struct ifnet
pointer locklessly and then pick the if_inet6 pointer from it and
dereference it. While the struct ifnet is freed via epoch_call(9), the
struct in6_ifextra until this change was not. For the forwarded packets,
or locally originated non-TCP packets we were probably safe due to the old
if_dead trick. But locally originated TCP packets may dereference
in6_ifextra via direct call into in6_selecthlim() from the tcp_output(),
before ip6_output().
NB: hypothetically a similar problem also applies to IPv4's if_inet pointer,
but there are no known panics, yet.
PR: 279653