Need one more syncthreads before syncgrid for Mixtral

Again not clear why any of this is necessary yet. One possibility is
that the writes are buffered or cached, and we could write directly to
L2 instead but that requires careful code modifications.

src/infer.cu

@@ -483,6 +483,7 @@ __global__ __launch_bounds__(1024, 1) static void kernel_forward(const __grid_co
 			}
 		}
 
+		__syncthreads(); // TODO: unclear why this is needed for determinism
 		syncgrid();
 		coopstage(args.perfstats, 0);