behavior discrepancy when aten::scatter.src receives a scalar index or scalar source

[linshokaku]

onnxscript/onnxscript/function_libs/torch_lib/ops/core.py

Lines 7739 to 7749 in 81f8444

	@torch_op(("aten::scatter.value", "aten::scatter.src"), trace_only=True)
	def aten_scatter(
	self: TReal,
	dim: int, # we have to use int here because ScatterElements() will use this attribute
	index: TInt,
	src: TReal,
	) -> TReal:
	"""scatter_add(Tensor self, int dim, Tensor index, Tensor src) -> Tensor"""
	update = op.Expand(src, op.Shape(index))
	return op.ScatterElements(self, index, update, axis=dim)

#2564 reports a similar issue, but I've also confirmed that aten::scatter performs implicit unsqueezing on 0-dimensional tensors, which requires correction.
When a 0-dimensional tensor is provided, ScatterElement will not function correctly.

>>> data = torch.zeros(3).float()
>>> torch.ops.aten.scatter.src(data, 0, torch.tensor(0).long(), torch.tensor(1).float())
tensor([1., 0., 0.])
>>> torch.ops.aten.scatter.src(data, 0, torch.tensor([0]).long(), torch.tensor(1).float())
tensor([1., 0., 0.])
>>> torch.ops.aten.scatter.src(data, 0, torch.tensor(0).long(), torch.tensor([1]).float())
tensor([1., 0., 0.])
>>> torch.ops.aten.scatter.src(data, 0, torch.tensor([0]).long(), torch.tensor([1]).float())
tensor([1., 0., 0.])
>>> torch.ops.aten.scatter.value(data, 0, torch.tensor(0).long(), 1.)
tensor([1., 0., 0.])
>>> torch.ops.aten.scatter.value(data, 0, torch.tensor([0]).long(), 1.)
tensor([1., 0., 0.])

[linshokaku]

This issue will be fixed by #2613.
May I request that it be reopened?