Remove barriers from sum factorization implementations

The y ar enot properly implemented right now, do not seem
to be necessary and conflict with recent changes in loopy.

python/dune/perftool/sumfact/realization.py

@@ -83,11 +83,6 @@ def _realize_sum_factorization_kernel(sf):
         ctags = ctags + ",vec"
         vec_shape = (sf.vector_width,)
 
-    # Put a barrier before the sumfactorization kernel
-    insn_dep = frozenset({barrier(depends_on=insn_dep,
-                                  within_inames=frozenset(sf.within_inames),
-                                  )})
-
     # Decide in which order we want to process directions in the
     # sumfactorization. A clever ordering can lead to a reduced
     # complexity. This will e.g. happen at faces where we only have