Move the CUDA implementation of atan2 (which was partially implemented in ATen) to ATen. (#26178)
Summary:
std::atan2 is not used because it does not work with HIP.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26178
Differential Revision: D17747897
Pulled By: VitalyFedyunin
fbshipit-source-id: b300f0573c431e1425644c9c1899d0b024c6a57c