Fix a deadlock happening if !NET_MPSAFE

If NET_MPSAFE isn't set, key_timehandler_work is executed with holding
softnet_lock. This means that localcount_drain can be called with holding
softnet_lock resulting in a deadlock that localcount_drain waits for packet
processing to release a reference to SP/SA while network processing is prevented
by softnet_lock.

Fix the deadlock by not taking softnet_lock in key_timehandler_work. It's okay
because IPsec is MP-safe even if !NET_MPSAFE. Note that the change also needs
to enable pserialize_perform because the IPsec code can be run in parallel now.

Reported by christos@

(ozaki-r)