intel: Limit TCS thread counts when barriers are in use
This fixes a bug that has existed in our tessellation support since the dawn of time. We've never really noticed it because, with only single patch mode, barriers are actually fairly uncommon because we can just delete them whenever instances <= 8. Even with barriers single patch still has far fewer threads involved in the barrier so the chances of it going south are smaller. When implementing 8-patch, however, this suddenly became a very visible issue.