[pytorch] refine dispatch keys in native_functions.yaml (2/N) (#45284)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/45284
This is the 2nd batch of the change described in #45010.
In this batch we relaxed some filters to cover more 'backend specific' ops:
* ops that not call any 'Tensor::is_xxx()' method OR only call
'Tensor::is_cuda()' - we are adding CUDA dispatch key anyway;
* ops that call other ATen ops but ARE differentiable - differentiability
is a fuzzy indicator of not being 'composite';
Inherited other filters from the 1st batch:
* These ops don't already have dispatch section in native_functions.yaml;
* These ops call one or more DispatchStub (thus "backend specific");
Differential Revision: D23909901
Test Plan: Imported from OSS
Reviewed By: ailzhang
Pulled By: ljk53
fbshipit-source-id: 3b31e176324b6ac814acee0b0f80d18443bd81a1