That way, we use the CudaNdarray, instead of transferring back and forth
to the GPU, and the Gpu ops have CudaNdarrays as inputs, not numpy
ndarrays or scalars.

This removes the need for a special 'ndarray' case, and should generalize to
Python and numpy numeric types, as well as CudaNdarray, and even other types.

Also test it.
This should fix the problem experienced by Justin Bayer.

Also remove the test that used them.

We use the ShapeFeature and SpecifyShape, now,
tag.shape is being deprecated.

Thanks Fred for the review.

* compute_test_value=True enables on-the-fly execution of perform method
* the user provides default test values through the tag scratchpad, i.e.
T.dmatrix().tag.test_value = numpy.random.rand(4,6)

Also remove the test that used them.

We use the ShapeFeature and SpecifyShape, now,
tag.shape is being deprecated.

Thanks Fred for the review.