[Bug Report] We should never import torch outside of using golden functions

[eyonland]

Describe the bug
We should never import torch outside of golden functions.

https://github.com/tenstorrent/tt-metal/blob/main/ttnn/ttnn/operations/binary_backward.py#L15

File "/home/blozano/anaconda3/envs/py310/lib/python3.10/site-packages/ttnn/operations/binary_backward.py", line 8, in <module>
    from ttnn.operations.complex_binary_backward import (
  File "/home/blozano/anaconda3/envs/py310/lib/python3.10/site-packages/ttnn/operations/complex_binary_backward.py", line 6, in <module>
    import torch

To Reproduce
Steps to reproduce the behavior:
Just import ttnn without torch and you will see this.

[patrickroberts]

How are the top-level statements like this supposed to reference torch then?

ttnn.attach_golden_function(
    ttnn.sub_bw,
    golden_function=lambda grad, a, b, *args, **kwargs: _golden_function_backward(
        torch.sub, grad, a, b, *args, **kwargs
    ),
)

[eyonland]

https://github.com/tenstorrent/tt-metal/blob/main/ttnn/ttnn/operations/binary_backward.py#L23

def _golden_function_backward(torch_op, grad_tensor, input_tensor_a, input_tensor_b, *args, **kwargs):
    import torch

    if torch.is_complex(input_tensor_a):
        if torch_op == torch.add:
            alpha = kwargs.pop("alpha")

[patrickroberts]

Okay, so don't use lambdas then. Gotcha.