Add FAQ-style QA

[tholor]

Addressing #42 and adding support for FAQ-style question answering.

Motivation:
In many cases, users already have a list of frequently asked questions (+ answers). Using this in combination with an extractive QA model can be very powerful as:

• FAQs are highly curated content (= very meaningful answers)
• Matching query with existing questions is computationally cheap and fast
• FAQs (should) capture the most common queries and can therefore significantly reduce the workload on the GPU

Let's implement a retriever that supports this via a similarity measure of embeddings. While the query embedding needs to be computed at inference time, we can precompute and index the ones from the FAQs. The model could e.g. be Sentence-BERT or a model trained on the Quora duplicate questions dataset.