Better ringbuffer-full handling

Currently if userspace manages to fill up the ring faster than the GPU can consume we (a) spin for up to 1sec, and then (b) overwrite the ringbuffer contents from previous submits that the GPU is still busy executing. Which predictably goes rather badly.

We could possibly just block forever until there is space or the GPU hangs? But some care needs to be taken as far as locks that are held while writing into the RB.

---

See: https://patchwork.freedesktop.org/series/89598/

I've picked up 2/2.. 1/2 needs to be re-worked now that we are using drm/scheduler.