Added optimization of useless elemwise node. Refactored those optimization and test to put them together.

 reverted the many copies scan used to do since it introduced bugs and the corner case it was suppose to address was fixed by Pascal L.

fix bug in conv op on the gpu that make the code crash in some case. We don't generate bad result, but read memory that was not usefull and outside the image range.

add new version of Windows documentation that use pythonxy to have a simpler installation. Their is test instruction for the gpu on windows.

make GpuFromHost return a dtype of the same dtype as the input. As CudaNdarray support only float32, an error will be raised.

fix MakeVector.make_node to upcast its input correctly. fix stack to don't downcast float* to int64 and to don't upcast inputs more then necessary.