scatter's backward formula should call scatter instead of scatter_

[zou3519]

🚀 Feature

Replace scatter_ in scatter's backward formula with scatter.

Motivation

functorch has a hard time with in-place operations and I don't see a good reason why scatter_ gets used here (please correct me if I'm wrong!). cc @albanD @soulitzer for sanity check - please let me know if I missed something here

pytorch/tools/autograd/derivatives.yaml

Lines 1187 to 1194 in 495e7e4

	- name: scatter.src(Tensor self, int dim, Tensor index, Tensor src) -> Tensor
	self: grad.clone().scatter_(dim, index, 0)
	index: non_differentiable
	src: grad.gather(dim, index)
	- name: scatter.value(Tensor self, int dim, Tensor index, Scalar value) -> Tensor
	self: grad.clone().scatter_(dim, index, 0)
	index: non_differentiable

[soulitzer]

Could it be because of this?
#56975

[ngimel]

#56975 should not affect this, as we can specify the argument with value, not src. @zou3519 yes, we can use scatter instead of scatter_, and avoid clone in this case, because it will be done in the scatter itself.