* TGI adapter and some refactoring of other inference adapters * Use the lower-level `generate_stream()` method for correct tool calling --------- Co-authored-by: Ashwin Bharambe <ashwin@meta.com>
