Noticed while working on #65.

The priority queue ram_duration/1 implementation takes the sum of all the ram durations of the sub-queues. That means, if a single sub-queue has a ram duration of infinity (which is very likely!) then the entire queue does.

I didn't really think that through. The point of what we're trying to do here is identify fast-moving queues so they can use more ram. Ideally each sub-queue would have its own ram duration, but in the absence of that we should probably take the minimum of the sub-queue ram durations, not the sum of them.