ac,radeonsi: cleanups, fixes, perf changes maybe
Mostly cleanups.
Broken gfx6 and gfx7 is fixed in the 2nd commit. SQ perf counters on gfx10.x should also work now. There are also other bug fixes.
Perf changes:
- TMPRING_SIZE adjustments
- LLVM 14 doesn't need the discard bug workaround (it should use Wave32 PS more often now)
- The tess factor ring size increased