Use Scan.L_op instead of Scan.grad() to help speed up the second derivative

Author: nouiz

theano/scan_module/scan_op.py

@@ -1931,8 +1931,7 @@ def get_oinp_iinp_iout_oout_mappings(self):
         return mappings
 
     # GRAD FUNCTION
-    def grad(self, inputs, dC_douts):
-        outs = self(*inputs)
+    def L_op(self, inputs, outs, dC_douts):
         if not isinstance(outs, (list, tuple)):
             outs = [outs]
         # `grad_step` equals the number of steps the original scan node has