feat: add `LLMReranker`

[micpst]

Feature description

Add a new reranker that will use LLM to rerank elements. There is a nice implementation by OpenAI, this cookbook can be used as a referance - https://cookbook.openai.com/examples/search_reranking_with_cross-encoders

Motivation

Most RAGs require some sort of reranking, setting up external provider in most cases is impossible due to clients constraints and setting up local model requires more effort than just adding a new LLM call. This reranker should speed up development process since some projects are already using similar implementations based on ragbits.

Additional context

No response

[micpst]

There may be a need to parameterize the logit_bias for the LLM call, which requires the caller to know the token ids for “yes” and “no” tokens for the specific LLM used. We can get this from litellm - https://docs.litellm.ai/docs/completion/token_usage (there is no public API for querying the tokenizer dict, but there must be something hidden inside the lib). Alternatively, the user should be able to provide these token ids themselves.