Fix for PyTorch 2.x: Remove _optimize_trace call & fix torch_node attribute access

[alz95]

Summary:
Recent versions of PyTorch (2.x) have removed the internal function torch.onnx._optimize_trace, causing an AttributeError in hiddenlayer. Additionally, subscript usage like torch_node[k] is no longer valid on torch._C.Node, resulting in a “TypeError: 'torch._C.Node' object is not subscriptable”.

This PR removes the _optimize_trace call in pytorch_builder.py and replaces torch_node[k] with the relevant attribute accessor methods (torch_node.f(k), torch_node.i(k), etc.). These changes allow hiddenlayer to generate graphs without error on PyTorch 2.x and Python 3.12.

Changes in this PR:

• Comment out torch.onnx._optimize_trace(...).
• Introduce a small loop to handle each attribute type in import_graph() (e.g. f, i, fs, is, s).
• Keep the rest of the logic (e.g. shape inference) intact.

Testing:

• Verified locally on PyTorch 2.0.1 / Python 3.12. The modified code no longer throws AttributeError or TypeError.
• Should maintain compatibility with older PyTorch versions as well, though not extensively tested.

Notes:

• This is intended as a minimal fix. If there are other parts of hiddenlayer that rely on removed/changed PyTorch APIs, further patches may be needed.
• Please let me know if you have any questions or suggestions.