have get_constant_value handle the case of Subtensor(Shape) when their is some dimensions broadcastable.

This should fix the bug in the buildbot about the grad of the neibhbors op.

Renamed temporary variable s_input into tmp_s_input to avoid confusion with s_inputs and thus make code easier to read

This fix the remaining case error in the buildbot. This don't hide the new case fused in the previous commit as their was error before it in the gpu 3d case.

The error in 4d are not real error as they are caused by limit in the number of input in a gpu elemwise op we can put in the gpu code that exist now.

Make the fusion of elemwise respect in all case the maximum number of op and make it fuse more element to get closer to it in some case.

This make the new failure in the Buildbot fail in less case.

remove a test that fail for not implemented feature as the original case is now tested in other tests

fixed optimization bugs triggered by propagation of shape 1 to unbroadcastable dims - hint: prefer broadcast_like() to alloc()

Updated Windows install instructions to say GPU does not work, also replaced refs to Python 2.7 by Python 2.6 because Python(x,y) is currently using 2.6

Fixed check of the existence of the cuda_ndarray module under Windows, where the extension is not .so

Fixed a bug reported by cityhall consisting in not calling tensor.as_tensor on arguments before calling make_node