Added proof of concept for torch.fx to MDF

[parikshit14]

Proof of concept for torch.fx to MDF #431

• Created pytorch_fx_to_mdf() which takes a pytorch model, argument as its input and converts to MDF model by node then edges then appends them to the current MDF Graph
• Currently works for examples/PyTorch/simple_pytorch_to_mdf.py https://github.com/ModECI/MDF/blob/main/examples/PyTorch/simple_pytorch_to_mdf.py and added support to use the similar approach
• In the conversion, I have still used ONNX in functions so that we can execute the MDF graph to verify the results but in future, we will need changes in execution_engine or maybe use another representation. (Didn't explore actr, ddm present in execution engine)
• Because of using ONNX functions, parameter.args dictionary is restricted to particular values of key for the respective functions.
• Tried to make it as general as possible and use the standard way as in pytorch_to_mdf(), typically the onnx::Reshape op requires further generalization
• Minor Fix for ORT 1.9 issue src/modeci_mdf/functions/onnx.py

ValueError: This ORT build has ['AzureExecutionProvider', 'CPUExecutionProvider'] enabled.
Since ORT 1.9, you are required to explicitly set the providers parameter when instantiating InferenceSession. For 
example, onnxruntime.InferenceSession(..., providers=['AzureExecutionProvider', 'CPUExecutionProvider'], ...)

@davidt0x @pgleeson

[davidt0x]

I think we need a more general way to identify the appropriate ONNX function for the torch op. I think this might be something that the torch and ONNX folks are working on as well, see here. It looks like they are developing a converter between torch fx IR to onnxscript. I didn't know that ONNX now has a thing called ONNX script. Allowing for composing ONNX ops in Python. Might be useful for re-writing the execution engine in fact.

Anyway, I think this issue needs to be fixed before merging.

[parikshit14]

I did go through their implementation and they seem to get the appropriate function from here. But cannot figure out how to use a key in _OP_OVERLOAD_TO_EXPORTER_KEY_TABLE dictionary to get the value.