[torchlib] torch.ops.aten.im2col

[BowenBao]

Since torchlib does not have an implementation for this op, the op is being decomposed by dynamo onnx exporter during export. There is a performance concern regarding unnecessary upcasting in the aforementioned decomposition, discussed more in pytorch/pytorch#121363.

On the other hand, the decomposition emits aten::index, which is exported as onnx::GatherND. Within the context of im2col, it is possible to instead export using onnx::Gather. It is unclear if the latter approach would yield higher performance.

Depending on both above factors, it might become appealing to add the im2col implementation in torchlib. Marking as low priority for now and no immediate action required.