late remove ones

tinygrad/opt/kernel.py

@@ -451,6 +451,8 @@ class Kernel:
       ret = op.replace(src=tuple(fixup_ast(x) for x in op.src)) # noqa: F821
       if op.op in GroupOp.Buffer and op in self.bufs:
         st = self.sts[self.bufs.index(op)]
+        # late remove all ones
+        st = st.reshape(tuple([x for x in st.shape if resolve(x != 1)]))
         # NOTE: if CONST got masked after applying opts, we create a new VALID
         if op.op is Ops.CONST and any(v.mask is not None for v in st.views): return op.view(st).valid()
         # otherwise we just replace the VIEW source