Unlike engine_start, which sets the engine's running field itself,
engine_stop instead requires its callers to clear the field. However,
not all of them do, and notably the timeout handlers do not, meaning
that after a request times out and we stop the engine we never start it
again, causing all future transfers on the channel to hit the software
timeout.

Instead, set running to 0 inside xdma_engine_stop to mirror engine_start
to both fix the bug and prevent future ones creeping in.