Add CUDA support for _dirichlet_grad (#21191)
Summary:
Changelog:
- Migrate _dirichlet_grad implementation from TH to ATen
- Add CUDA support for _dirichlet_grad
Closes #11030.
Closes #15773.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/21191
Differential Revision: D15660330
Pulled By: ezyang
fbshipit-source-id: c8ad5b80366e5348139ce9be10400f22fc430344
