[inductor] fix incorrect strides in copy() decomp, fix hf_LongFormer + hf_BigBird errors (#100115)
Fixes https://github.com/pytorch/pytorch/issues/100067 and https://github.com/pytorch/pytorch/issues/93428.
See the comment [here](https://github.com/pytorch/pytorch/issues/100067#issuecomment-1523856970) for details. The bug was that the decomposition that inductor uses for `aten.copy` doesn't respect the strides of the input in all cases. The fixes that I added should work, but will be pretty slow - we allocate a tensor (potentially larger than `self` if `self` is a slice), and perform an `as_strided_scatter` + `as_strided`. Longer term, stride-agnostic IR should let us remove this decomp? cc @soumith @voznesenskym @penguinwu @anijain2305 @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @Xia-Weiwen @wenzhe-nrv @jiayisunx @peterbell10 @desertfire @ngimel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100115
Approved by: https://github.com/albanD
