Add support for ONNX type and tag broadcasting

[RunDevelopment]

Closes #1370

This adds support for type and tag broadcasting for all ONNX models. While we did have type broadcasting for ONNX models before, those broadcasts did not contain much information. Now they contain the model's input/output channels and scale, just like for PyTorch and NCNN models. This finally allows us to compute output types for ONNX's Upscale Image node.

A model's input/output channels and scale(s) are detected using onnx.shape_inference [1] [2]. This allows us to statically determine the shape of the output tensor for a given input tensor shape without actually running the model (=no inference session). While this is roughly 10-100x faster than creating an ort.InferenceSession, it's still not free. Loading generic ONNX models will now take roughly 3-4x longer. This might sound bad, but it's not. E.g. an ESRGAN ONNX model previously took 0.11s and now it takes 0.38s. A lot slower, but not a problem.

---

Changes:

• Extract out loading ONNX models into its own file. This ensures that the type definitions for ONNX models don't depend on the ONNX package.
• Store ONNX model information in new OnnxInfo class.
• Removed is_rembg_model function, because it wasn't necessary.
• Refactored create_inference_session.