import the version of theano and nothing else. In some case we could import the version from another project.

the Theano type

refactored TensorSharedVariable test of shared options to be able to reused them for CudaNdarrayShared.

to check for input aliasing and removed the transitivity assumption on the
aliasing.

mutable or borrowed

Make PatternSub hangdling of allow_multiple_client with dictionnary that don't define it to default to the value of the parent nod instead of False.

Also fixed tensordot to accept numpy syntax for axes.

inplace computation

and inplace operations are used we are not getting the right results. Note
that is not easy to trigger this behaviour.