Support aten::split_with_sizes_copy.out/aten::_chunk_cat/aten::_chunk_cat.out to align with CUDA as a fast pass

[zhangxiaoli73]

🚀 The feature, motivation and pitch

Motivation

torch.split_with_sizes_copy.out is called in PyTorch FSDP2 https://github.com/pytorch/pytorch/blob/main/torch/distributed/fsdp/_fully_shard/_fsdp_collectives.py#L105

CUDA has a fast pass implementation for this aten op. https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/native_functions.yaml#L14757 . To achieve same performance as CUDA, please enable a similar fast path for XPU device.

Other cuda specific aten ops in FSDP2:
aten::_chunk_cat
https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/native_functions.yaml#L5720

aten::_chunk_cat.out
https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/native_functions.yaml#L5725

Customer impact
FSDP2 performance on XPU device.

Alternatives

No response

Additional context

No response

[xytintel]

#1089