panfrost: Faster fences

Given the kernel has to manage a lot of this anyway, we can simplify the fencing implementation and reduce CPU overhead by eliminating redundant syncobjs. Note this does not go as far as it 'could'. Indeed, the entire batch_fence struct could be rolled up into the batch, but I'd like to keep this MR focused on syncobjs for easier review.

I saw about a 10% reduction in frametime in an apitrace of chromium but testing is desired.

Cc @bbrezillon