`torch.add` and `torch.sub` return incorrect results when computing complex numbers containing `inf` on both CPU and GPU

[rookieLiu2018]

🐛 Describe the bug

When adding two complex128 tensors with infinite real parts (e.g. (inf+1j) + (inf+2j)), torch.add returns inf+nanj instead of the correct inf+3j.
When subtracting a complex128 tensor with an infinite imaginary part from another (e.g. (3+3j) - (4+infj)), torch.sub returns nan-inf j instead of the correct -1-inf j.
These errors occur identically on both CPU and CUDA.

import torch

# CPU
a = torch.tensor(complex(float('inf'),1), dtype=torch.complex128)
b = torch.tensor(complex(float('inf'),2), dtype=torch.complex128)
c = torch.add(a, b)
print("CPU add result:", c)

if torch.cuda.is_available():
    a_gpu = a.cuda()
    b_gpu = b.cuda()
    c_gpu = torch.add(a_gpu, b_gpu)
    print("GPU add result:", c_gpu)


# CPU
a = torch.tensor(complex(3, 3), dtype=torch.complex128)
b = torch.tensor(complex(4, float('inf')),   dtype=torch.complex128)
c = torch.sub(a, b)
print("CPU sub result:", c)

if torch.cuda.is_available():
    a_gpu = a.cuda()
    b_gpu = b.cuda()
    c_gpu = torch.sub(a_gpu, b_gpu)
    print("GPU sub result:", c_gpu)

Output:

CPU add result: tensor(inf+nanj, dtype=torch.complex128)
GPU add result: tensor(inf+nanj, device='cuda:0', dtype=torch.complex128)
CPU sub result: tensor(nan-infj, dtype=torch.complex128)
GPU sub result: tensor(nan-infj, device='cuda:0', dtype=torch.complex128)

Versions

PyTorch version: 2.5.1


cc @ezyang @anjali411 @dylanbespalko @mruberry @nikitaved @amjames