Support DeepLab ONNX model as exported by PyTorch's dynamo-based exporter

[robertknight]

Changing the DeepLab ONNX export script to use the new dynamo-based exporter, by adding dynamo=True to the torch.export call, produces an ONNX model which rten-convert cannot convert. This issue tracks the various features which are missing to make it work.

• Support for Constant operators which store their value in an attribute other than value (eg. value_ints, value_floats). See spec (Support value_{int, ints, float, floats} attributes in Constant op #649)
• Support CastLike operator (Implement CastLike operator #646)
• Support for start and end attributes in Shape operator (Support start and end attributes for Shape operator #648)
• Support for pytorch_half_pixel value for coordinate_transformation_mode attribute of Resize operator (Support pytorch_half_pixel value for coordinate_transform_mode #647)
• Support for function operators (eg. _aten_native_batch_norm_inference_onnx in this model). Update: This requirement goes away if the model is optimized after export. See Add option to use dynamo ONNX exporter in export-timm-model.py #651. After fixing that a subsequent issue arises where inference produces different results than for TorchScript.
• Investigate different inference result when comparing optimized Dynamo-exported model against TorchScript-exported model (Fix pytorch_half_pixel resize mode in Resize op #654)

[robertknight]

I picked DeepLab for testing Dynamo-exported models because there is a convenient script in rten-examples/src/export_deeplab.py for exporting. Exporting other models with the Dynamo exporter will likely encounter additional issues.

[robertknight]

From testing a few timm models exported with tools/export-timm-model.py the most common issues with other vision models are:

• Lack of support for functions (This usually becomes a non-issue if the ONNX model is optimized before being saved, as functions get inlined.)
• Lack of support for Dropout operator. This is often a no-op node since the dropout ratio is zero. Optimizing the ONNX model should remove this operator, but it makes sense to support it anyway. (Resolved via Implement Dropout operator #652)

[robertknight]

As of #654 it is possible to export the DeepLab v3 model using the TorchDynamo exporter, convert it and then run it.