Change __CUDACC__ and __HIPCC__ to __CUDA_ARCH__ and __HIP_ARCH__ in NumericUtils.h (#39213)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/39213
This PR fixes the problem that [__expf/__logf/__tanf](https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH__INTRINSIC__SINGLE.html) are "intrinsic functions that are only supported in device code", so nvcc doesn't recognize them if it compiles host code. So `__CUDACC__ ` should be replaced with `__CUDA_ARCH__ `
Test Plan: Imported from OSS
Differential Revision: D21779132
Pulled By: pbelevich
fbshipit-source-id: b326e2135525b6a1f2392f8d1c17b735d8ef431a