[ATen core IR] Register additional ATen operators as core (#110882)
## Context
For more context, please refer to [this PyTorch forums post](https://dev-discuss.pytorch.org/t/defining-the-core-aten-opset/1464).
This PR registers some additional ATen operators as `core`, based on feedback from the forums post as well as the experiences from adding other core ATen decompositions.
The ATen operators registered as core in this diff, with the associated reasoning, are:
ATen op | reasoning
--|--
aten::atan2 | This operator often maps to a hardware intrinsic.
aten::diagonal | There is no straightforward decomposition for this operator.
aten::empty_like | Decomposition for this operator would require `as_strided` to retain the strides of the input tensor, which should be avoided.
aten::expm1 | This operator often maps to a hardware intrinsic; Furthermore, decomposing it will negatively impact the numerical precision of the output.
aten::full_like | Decomposition for this operator would require `as_strided` to retain the strides of the input tensor, which should be avoided.
aten::log10 | This operator often maps to a hardware intrinsic; Furthermore, decomposing it will negatively impact the numerical precision of the output.
aten::log1p | This operator often maps to a hardware intrinsic; Furthermore, decomposing it will negatively impact the numerical precision of the output.
aten::log2 | This operator often maps to a hardware intrinsic; Furthermore, decomposing it will negatively impact the numerical precision of the output.
aten::pow.Scalar_Tensor | This is a Scalar variant of pow.Tensor_Tensor, which is a part of core.
aten::resize | There is no valid decomposition for this operator.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110882
Approved by: https://github.com/lezcano
