feat : better support for onnx graph import and inputs/outputs

Context

Add some corner test cases and fix ONNX model import for managing better:

• multiple reference of the same output name
• ordering of inputs and outputs, when some are ignored/unused/undefined they are ignored in the final Aidge ordered inputs/outputs list,
• forward references of output names, resolved to the first forward definition (allows cyclic graphs)
• multiple references to inputs, for which some identity nodes must be created

Actually the transformation to the input ONNX schedules data dependence graph, expected to be a topological order of nodes, to the Aidge data flow graph (static single definitions) is now done in a separate functions which computes and unambiguous DFG given the above choices.

Modified files

First commit:

• test_unit_cases_import.py/test_unit_cases_import_forward.py: new file adding some unit test cases for corner cases
• onnx_helpers.py: add utility function for creating ONNX graph from simple test cases described as dict
• test_import_export.py: minor changes

Second commit:

• onnx_import.py: add _get_dataflow_graph() function and rewrite import. All previously importable graphs are still importable.
• update test cases now passing

Detailed major modifications

The ONNX import now defined the ordered inputs, through set_ordered_inputs.

Add new tests with corner cases in new test files: test_unit_cases_import.py/test_unit_cases_import_forward.py

TODO

• Merge or iterate over MR !75 (merged) and MR !76 (merged) first