Add String data type support

[marshallpierce]

• Tensor construction now uses different logic for primitive types (using their native in-memory layout) and strings (filling the tensor with onnxruntime's FillStringTensor).
  • Extended TypeToTensorElementDataType to also be able to expose utf8 contents, if present
  • Utf8Data trait to make it possible to use both String and &str
• call_ort helper that takes care of mapping the ort status to a Result so you can't forget to do it
• print_structure example that shows the inputs and outputs of an .onnx model (names, types, etc)

---

Without this, interacting with session inputs for a model with a string input would lead to SIGABRT as the auto-generated Debug, etc, didn't know what to do with an enum variant of 8 (onnxruntime's string type).

If this looks good, I'll work on an integration test with a trivial model to ensure that feeding data through actually does work, but I wanted to get feedback on the approach first.

[marshallpierce]

Strings outputs aren't properly handled, and also even a trivial string model that only applies Tensorflow's unique produces multiple outputs of different types, so output types will need to become more flexible. Returning to draft status until I have that resolved.

[nbigaouette]

Thanks for your work and sorry for my late reply!

Except from small nitpicks, I wouldn't mind merging this. I would even like to see a small/simple .onnx model committed to test the functionality. Can you create a one that would take a <10s KB?

Thanks!

[marshallpierce]

I'll tidy up this stuff per your comments. I've found that just applying TensorFlow's unique as an ultra simple "model" produces a tiny 424 byte .onnx artifact that doesn't need any custom operators, so no problem checking that in, but it produces outputs of both string and int types. Could add a test that it doesn't crash when you run it with string inputs, but can't really meaningfully assert on the output until (1) string output and (2) dynamic type output are addressed in a follow up PR.

[nbigaouette]

Awesome! Thanks for this!! 👍