Even though all rendering should have been flushed at the end of the
previous requests, add an extra flush after switching to the
kernel_context. As the switch to the kernel_context is used when idling
the gpu (e.g. suspend), having an extra layer of paranoia to ensure
everything is flushed to memory seems sensible.