🐛 Describe the bug

There is perhaps a large discrepancy (requiring atol being increased to 3e-4) between the before+after #102427 implementation of tan/tanh on CUDA with complex values.

Repro:

import torch
x = torch.tensor(-7.8167-0.0451j, device='cuda:0')
torch.set_printoptions(precision=10)
print(torch.tan(x))
print(torch._foreach_tan([x])[0])
print(torch._foreach_tan([x.to(device="cpu")])[0])

Before:

After:

I happened to observe this when debugging why some tan tests were failing when I added new sample inputs to the foreach tests in #109402 (comment).

Versions

trunk

cc @ptrblck @mruberry @parth-desai @peterbell10