Precompute ephemeral keypairs for ntor
In the ntor handshake, on the server side, we need to generate a new y,Y pair for every handshake. But there's no reason this needs to happen on demand; instead, it could be lovely to do batches of them ahead of time, so that we can respond to requests faster. This wouldn't lower our CPU requirements when we were under load, but it might help us with circuit creation latency under less-than-horrible loads.
Before we go ahead and do this, it might be clever to verify that there are in fact times when cpuworkers are idle on busy servers.