aka original sample prediction (or predict x0) E.g. useful for methods like self-attention guidance (see equation (2) in https://arxiv.org/pdf/2210.00939.pdf)