feat: Implement Comprehensive ONNX Export Testing and Validation (#654)

[SHIVANSH-ux-ys]

Description

This PR addresses #654 by implementing a robust validation rig for ONNX export. It moves the ONNX export functionality from "experimental" to a reliable feature by fixing several core bugs and adding comprehensive testing.

Key Changes:

• Core Logic Fixes: Updated save_model_for_inference to handle dynamic batch sizes and feature configurations (only continuous, only categorical, or mixed).
• State Management: Added explicit .eval() mode switching during export to ensure consistency.
• Dynamic Axes Bug: Fixed a bug where only categorical inputs had dynamic axes, which caused issues for continuous-only models.
• Dependency Formalization: Added onnx and onnxruntime to optional dependencies.
• Robust Testing: Fixed a NameError in the test suite and updated test_save_for_inference_onnx to perform numerical validation comparing PyTorch and ONNX outputs.

Verification:

• Verified logic for models with mixed, continuous-only, and categorical-only features.
• Added onnx.checker validation during the export process.

Closes #654