freedreno/ir3: sched sync improvements

Plus a crude heuristic to limit tex prefetch in short shaders. Together these help gfxbench fill2 significantly (gets us most of the way to blob scores, adding tiled etc2 on top gets us the rest of the way there).

Limiting the prefetch only effects two manhattan shaders, and framemetrics says the difference is in the noise. The changes to limit (sy) syncs in particular hurt register footprint in a handful of mh shaders, but again the counters say it is for the best. Overall, helps mh ~2-3%