MAINT: Remove unnecessary check from before swapaxes (API change)

Make GPU Ops accept TensorType output gradients (transfer them to GPU)

Fix opt error when trying to access a python list with a Type object.

Make test more resistent to different seed value.

Image2Neibs.perform

correctly fix the test with numpy 1.7.1

Tell witch version of NumPy have this bug.

Add MakeVector.c_code

In NumPy, I have a PR that make it use array_priority for comparison. The new test assert that it don't work, or work correctly.

Also add doc about witch linker support allow_gc.

fixes gh-1288

This check was a speed optimization in the numpy code. Due to
a small change in the MapIter API, the check is now wrong in
corner cases and becomes fully unnecessary.
The previous API had not been public in any release numpy
version.

Recent scan bugs

the unbroadcast(shape_padleft(x)) might not be of the same type as x,
specifically when x is CudaNdarrayTyepe,  unbroadcast(shape_padleft(x))
is a TensorType.