Symbolic shapes ignored when exporting to ONNX with Torch backend

[DLumi]

When trying to export model containing symbolic / non-fixed shapes, the exporter bakes in the lowest possible shape (i.e., 1).

Colab to reproduce:
https://colab.research.google.com/drive/1H81pAhOWr42YFnzrAT4zGI3XooXvM8oe

Suggestions:

1. model.export() already supports **kwargs, and torch specifies dynamic_shapes / dynamic_axes for torch.onnx.export(), so just expose those, and it would work fine as is
2. (better ux:) calculate dynamic_shapes / dynamic_axes automatically based off the input spec
3. or all of the above

[DLumi]

By the way, torch.onnx.export has dynamo=True by default, so I'm guessing, this line doesn't do what it's supposed to:

keras/keras/src/export/onnx.py

Lines 162 to 170 in 38687a6

	# Fall back to the TorchScript-based ONNX exporter.
	torch.onnx.export(
	model,
	sample_inputs,
	filepath,
	verbose=actual_verbose,
	opset_version=opset_version,
	input_names=input_names,
	)