Separate Indexing Logic from Inference

[jdchawla29]

For #30 I basically used Claude to refactor the classes, no breaking changes introduced to the existing api, RAGMultiModalModel will internally use IndexManager (can also be a wrapper over a VectorDB later on). All tests pass and the quick_overview.ipynb example works as earlier Let me know if any changes are needed.

Also an aside: I used Byaldi locally on a MacbookPro M3 and I was surprised that it's just as fast as it would be on google collab.