(when TorchBind custom classes work with PT2). There are a number of assumptions that we should check:

1. these ops don't support autograd
2. we don't end up putting a Tensor inside of a custom class object and then mutating it.
3. the abstract impls match eager mode

cc @mruberry @ZainRizvi @ezyang @chauhang @penguinwu @msaroufim @bdhirsh @anijain2305