Added a class for ONNX models

[RunDevelopment]

Changes:

• ONNX models are now wrapped in a class that only contains the model bytes.
• The inference session is now lazily created using the get_onnx_session function. This function will create a new session if no existing session is associated with the model.
• Fixed that ConvertOnnxToNcnnNode still used the old type and crashed at runtime.

The divide between the model (bytes) and the inference session is necessary because the model class defines an interface that has to be usable even if the ONNX runtime package isn't present and no inference session can be created.