feat: support prompt caching and apply Anthropic's long-context prompt format

[undo76]

Several improvements:

• We return the document id in the context to make it available to the model. chunk_id and document_id not accessible #50
• Better reconstruction of segments considering headings hierarchy
• Fix order of segments. We use the sum of the reciprocal ranks of the chunks by default, while allowing passing a scoring function.
• Now we write the context in the user prompt, but we can configure to pass it in the system prompt or in a separate system prompt. Order of the system message may prevent caching #49

[lsorber]

Thanks for the PR. It's a fairly big one so I have some more comments than the previous one 😅.

[undo76]

Great changes, I think it is a big improvement. Thanks!