Remove conj kernels for real dtypes (#80374)
`conj_physical_stub` is currently implemented for all dtypes despite
it just being a plain copy for real dtypes. So, instead we should
defer to the existing copy kernel in these cases.
On my build for one CUDA architecture, I see a 2.2 MB decrease in
`libtorch_cuda.so` size.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/80374
Approved by: https://github.com/ngimel