// Tensor or IndexedSlices or None, which here we map to nullptr. Gradients need
// to allow their size to be computed and they need to be passable to a backward
// function and deleted (as the backprop code creates lots of gradients the user
// is not interested in).
//
// BackwardFunction needs to be a closure which stores intermediate activations
// from the forward computation and calls a vector-jacobian product function