Refactor output processing in OnnxInferenceModel prediction methods

[juliabeliaeva]

This PR adds:

1. Function OnnxInferenceModel.predictRaw is introduced to be used for custom output processing and in other prediction methods to remove duplication between them.
2. Utility functions to extract prediction result without casts from the OrtSession.Result object. There are separate functions for each buffer type, unfortunately did not find a way to implement them without copy-paste.

This allows to get rid of predictRawWithShapes method and replace it with the more customizable alternative. Also we can now pass OrtSession.Result directly to the OnnxHighLevelModel#convert function.

Fixes #358