intel/compiler: fix URB fences for TASK/MESH

This is an attempt at fixing the problem that's worked around in !16197 (merged) and reproducible by mslusarz/crucible@d95e5858. Unfortunately, these changes are not enough to fix the issue but are still something we should land.

intel/compiler: emit URB fences for TASK/MESH - This is something I noticed when I inserted scoped_barrier at NIR level, but later decided against it and implemented necessary barrier in intel/compiler: insert URB fence before task/mesh termination.

intel/compiler: there are 4 types of fences on gfx >= 12.5 doesn't fix any known bug. I stumbled on it when I inserted an overly broad barrier.