Added QA handler using raw logits to return best answer span instead of pipeline()
This handler processes question/context inputs using the model’s start and end logits to score all valid spans (up to max length) and return the most confident one.
Avoids pipeline()’s greedy argmax approach, which often misses better spans due to oversimplified selection.
While slightly more compute-intensive (O(n × max_len) vs O(n)), this method offers much higher accuracy and control for real-world QA systems.
