Support for ONNX function import by converting functions to modules in Burn

[antimora]

Description

Add the capability to import ONNX functions into the Burn deep learning framework by converting them into modules. This feature aims to handle ONNX functions, which essentially serve as mini computational graphs, containing nodes, inputs, and outputs.

Motivation

1. Seamless Interoperability: Many deep learning practitioners work with PyTorch, which can export models that maintain their module structures via functions. Achieving compatibility between PyTorch-exported models and Burn would facilitate a smoother migration and collaborative workflow.

2. Model Integrity: Preserving the module structure during export/import processes ensures that the model's architecture remains intact, which is crucial for debugging, modification, and further development.

3. Reuse and Modularity: Converting ONNX functions to modules in Burn can also facilitate model reuse and modularity, making it easier to plug in or modify specific parts of a model.

4. Streamlined Development Cycle: This feature can simplify the model import process, saving time and computational resources, especially when transferring complex models between different frameworks.