Improvements to code exporting

[kevinwuTT]

Original PR: #611

Improvements

• Analyze and report accuracy failures in Github CI
  Currently, the standalone scripts will run and fail silently. The logs will show which models fail and where, but this is not easy to view.
  Possibly dump a json file if a comparison fails and have the CI process it.

• Organize into a base module that only exports standalone runnable code.
  Separate modules to export additional code:

  • Compare accuracy
  • Profiling helper functions for tracy
    https://github.com/tenstorrent/pytorch2.0_ttnn/tree/kw/gen_acc_tests_for_profiling

• Separate forward functions into separate definitions that mimics the original test model more closely.
  It's currently simpler to fuse all of the forward functions since the input data is only applicable for the first forward function. But this may not be representative of the original model with separate graph breaks.

• Using safetensors instead of pickle
  There are issues with saving the tensors in this format because of share memory. save_model won't work because this is not a torch.nn.Module.

  RuntimeError: 
      Some tensors share memory, this will lead to duplicate memory on disk and potential differences when loading them again: [{'arg0_1', 'arg27_1'}].
      A potential way to correctly save your model is to use `save_model`.
      More information at https://huggingface.co/docs/safetensors/torch_shared_tensors
  

• inspect.getsource() assumes the target function is always accessible and valid. This may fail if node.target is dynamically created or a non-source-mapped function.