This is a second type of use for the scratchpad: a single master
thread is responsible for launching all data movements, but the tiles
are worked on in parallel. We support this model by using a sequential
scratch on top of a parallel dma.