Use DMA engine for large memory copies

We currently use a single global queue, but large memory transfers should probably use a special queue with FLAG_COPY set so that the DMA copy engines can be used. We'll probably need to order operations on that queue wrt. to the global queue (using events?).

Start by tracing the current single global queue and the handling of large memory transfers. Investigate how a separate queue with FLAG_COPY and event-based ordering could coordinate with the global queue. Done means large copies use the DMA copy engines without violating operation ordering.

Written by the indexing model from the issue text.