This PR unwrap undefined tensor when interfacing with the rest of the pytorch system.
This is because undefined is handle specially in other places.
For example, in autograd, a undefined value will get ignored, but a checkpointed-undefined value will try to be process by the system, leading to an error (because there is nothing to back propagate against).
This happen when you try to do a second order gradient over convolution2d, because conv2d return undefined for values when it dont need to to save time.

There is also some minor fixes that made us play better with autograd. pytorch has weird and poorly documented semantic regarding requrires_grad and dispatch key that I didnt got right.