Research spill prefetch when a low priority task may raise in priority (retry framework)

[abellina]

Related to #7710 and #7253

As threads up in priority and get more of the GPU, should we also move spilled buffers for such high priority thread(s) into GPU memory? The idea is to figure out whether we want some smart "prefetch" to occur.

I can see a number of issues with this since we do not know what pattern the high priority thread wants to follow with the spilled buffers. If we have cases where we always need N spilled buffers before we can call into cuDF then that seems somewhat reasonable, and so this issue could evolve into creating a batch unspill API.

Tracking this to check once things have settled a bit if there are cases where high priority tasks are stuck waiting or iterating back and forth with the spill framework.