Refactor OnnxInferenceModel to reduce code duplication

[ermolenkodev]

I think it would be a good idea to unify this (predictSoftly) method with predictRaw (as a separate PR).

Originally posted by @juliabeliaeva in #356 (comment)

[zaleslaw]

Keep in mind the predictSoftly method for the TensorFlow models before refactoring

The initial idea of the predictSoftly method is the following: return the vector of probabilities instead of the class label calculated via argmax on the vector of probabilities.

The semantic for the predictRaw is more common – to return just low-level output with a few tensors, for example, not just 1D vector