Improve scheduler's wakeup performance

[nyh]

For N cpus, the scheduler uses N^2 wakeup queues, where each CPU tells another CPU which thread it wants to wake.

We can improve performance of these wakeup queues in two ways - 1. replace the queue implementation but a faster one, and 2. avoid unnecessary IPIs

Some more details:

1. The current implementation uses a linked list. This implementation allows multiple producers, a feature we don't need here. Rather, we can use a ring-based single-producer single-consumer queue (see lockfree::ring). We measured the ring implementation to be 3 times faster than the linked list one. Note that the ring is limited in size, so if it gets full we need a fallback - which should probably be the existing linked list code.
2. Only send an IPI when adding an item to an empty queue.
   Can we avoid IPIs in more ways?

[nyh]

The two things described in this issue are basically done, so closing this issue:

Commit bca7bbb reimplemented the wakeup queues. We still have N^2 of them, but each is really optimized for single-producer, and as explained in that commit, context-switch performance did get better.

Commit 5b05bad added a bitmask which says that one of the queues is non-empty, and when this bitmask is nonzero, we can avoid an IPI.