Do the vector workaround only on TensorVariable and document it.

theano/scan_module/scan_op.py

@@ -422,9 +422,15 @@ class Scan(PureOp):
                 raise ValueError('For output %s you need to provide a '
                                  'scalar int !', str(outer_nitsot))
         assert len(new_inputs) == len(inputs)
-        self.vector_seqs = [seq.ndim == 1 for seq in
-                             new_inputs[1:1 + self.n_seqs]]
-        self.vector_outs = [arg.ndim == 1 for arg in
+
+        # The vector_seqs and vector_outs are just a workaround
+        # strange NumPy behavior: vector_ndarray[int] return a NumPy
+        # scalar and not a NumPy ndarray of 0 dimensions.
+        self.vector_seqs = [isinstance(seq, tensor.TensorVariable) and
+                            seq.ndim == 1 for seq in
+                            new_inputs[1:1 + self.n_seqs]]
+        self.vector_outs = [isinstance(arg, tensor.TensorVariable) and
+                            arg.ndim == 1 for arg in
                              new_inputs[1 + self.n_seqs: (1 + self.n_seqs +
                                                     self.n_outs)]]
         self.vector_outs += [False] * self.n_nit_sot