Returning Confidence Score For Extractive QA Task When Using Non-Pipeline Approach

# 🚀 Feature request

HF's Extract QA pipeline provides an excellent interface for start. It returns 4 values including a **probability score / confidence score**. Unfortunately same is not the case when using the non-pipeline approach i.e using model and tokenizer for question answering.
[Both methods are mentioned here, The pipeline one and the other](https://huggingface.co/transformers/task_summary.html#extractive-question-answering)

## Motivation
The confidence score will help a lot in various tasks. For example when I am developing a complete pipeline for QA, consisting of recall, retriever and some other models for entity matching and etc. I need to calculate scores of each models and then rank the final list of documents based on the weighted sum of score from each model. I believe this is a very common practice among NLP practitioners and not just for QA task. The point is confidence scores are usually a pretty standard requirement for each model output because we have to take further actions based on its score.

## Your contribution
I want to. but unfortunately I am not at the level where I can understand the code. I have went through the code and I believe its the "decode" function in "QuestionAnsweringPipeline" class which has the code the which generates the probability scores. If you guys can just provide an interface for it or provide docs for how to calculate this score using the model and tokenizer approach then that would be great too. And if you do decide to do this then please also add this addition to docs in the link mentioned at the top.

Thanks.