Port `sort` to structured kernels.
Tracking Issue: #55070
This PR relands #67016, with the modifications discussed in https://github.com/pytorch/pytorch/pull/67015#issuecomment-982004500.
In summary, we call `infer_dense_strides` on the input's strides, and pass it to `set_output`. Meaning that if one of the outputs is resized (by a `resize_output` call), we will also restride such an output using the dense strides of the input.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/72058
Approved by: https://github.com/bdhirsh
