Add modelwrapper methods to get network global input and output tensor names

[rothej]

This fix adds the following helper methods in ModelWrapper:

global_in = model.get_first_global_in()
global_out = model.get_first_global_out()

This is meant to handle the following FINN issue: Xilinx/finn#1494

Once this PR is approved, can use the new commit hash in the FINN fix and remove the duplicate definitions from my FINN fix.

[rothej]

Want to make a note that all tests pass if I do NOT use multiple cores in parallel. Tests fail for both main branch and my branch (for tests unrelated to my fix) if I use pytest -n auto --verbose. All tests run fine if I don't use parallelism (pytest -p no:randomly and pytest --verbose both pass). I assume this is something the maintainers are aware of, but wanted to mention just in case.

[iksnagreb]

To avoid potential confusion: As ONNX models in general can have multiple inputs and outputs, would it make sense to explicitly name these methods get_first_global_in and get_first_global_out, as you are singling out the first one as special? Or what about get_first_input_name/get_first_output_name to reflect that this is referring to the tensor name and not the tensor itself?

[rothej]

I'm okay with whatever naming conventions make sense

@auphelia any input? I figure these'd have to match on the finn side as well

[auphelia]

Thanks for looping me in! I agree with @iksnagreb, making it explicit that this returns the first input and first output would help avoid confusion. Even if most FINN test cases currently only use a single input/output, it’s good to keep the naming clear, to avoid confusion in the future.
From my side, the proposed clarification sounds reasonable and consistent with how the function is used in @rothej's corresponding FINN PR. And yes, @rothej, you are right, the FINN PR would then need to be adjusted to use the new naming.

[rothej]

@auphelia @iksnagreb names fixed. pytest passes as before