Fully support index-free encoding and querying

[bclavie]

Most of the necessary functions are currently present, but not fully implemented.

While it can run slower and is memory intensive, there's nothing stopping us from querying smaller collections on-device, by encoding the documents and performing the computation without building an index.

The goal here would for RAGPretrainedModel/the ColBERT model class to support an additional .index_free_encode() and .index_free_search() (or better naming) functions. The former would encode docs and store their representation in-memory, while the latter would query it.

Functionally very similar to rerank(), except encoding & searching are performed at different stages and the encodings are stored, rather than on-the-fly for rerank.