Keep track of kswapds: it's OK that they get moved off a node when the
last CPU goes down, but when a CPU comes back, we should try to move
the kswapd back onto its node.