make theano.grad() build a GradOp and unfold it during compilation

[nouiz]

This would help some usecases like:

• Taking the grad on an unmerged graph.
  • Optimizing it well in all cases would request the lift of the elemwise add on all existing op. Which is prohibitive.
  • unmerged graph happen sometimes with clones when people try to make multiple replacements.
• We have opt to take the unfolded softmax graph and put the op. But with this change, we won't need an opt to recognize the unfolded grad of softmax to the softmaxgrad (we don't have it now). So this would request less optimization.