Add some TF ops and add XLA backend

[Co1lin]

No description provided.

[ganler]

Thanks for the contribution!!! In addition to those minor issues, would you also help add a dtype_test for tensorflow? Like https://github.com/ise-uiuc/nnsmith/blob/main/tests/torch/test_trt_backend.py#L56

[ganler]

Would you mind if I prioritize merging #41 as it alignes with an alpha release? I can help you sync this PR with latest branch.

[Co1lin]

Would you mind if I prioritize merging #41 as it alignes with an alpha release? I can help you sync this PR with latest branch.

Sure. You can merge that, and I will sync with the latest main branch.

[ganler]

The conv2d spec should be fixed. TF's padding mechanism is quite different from torch. See https://www.tensorflow.org/api_docs/python/tf/nn#notes_on_padding_2

[ganler]

There are still a few things to fix later:

• reproducibility: the seed obtained from the fuzzing loop cannot reproduce the same model in nnsmith.model_gen.
  • Oh I figured out: We need to both leverage the (model, backend) topset to have the exact same operator candidates.
• Even the graph mode of TensorFlow might not support enough Data Types -- we also need to do dtype testing for the model generator.