QA Evaluation: Exact Match and Token F1

For question answering and RAG, answers rarely match the gold answer character for character. The SQuAD metrics normalise text first:

1. lowercase,
2. remove punctuation,
3. remove the articles a, an and the as whole words,
4. collapse whitespace.

Then:

• Exact match: 1.0 if the normalised strings are equal, else 0.0.
• F1: token-level overlap (counting repeated tokens with a multiset), $F1 = \frac{2PR}{P+R}$. If either side has no tokens, F1 is 1.0 when both are empty and 0.0 otherwise.

Implement normalize_answer, exact_match and f1_score, then score(prediction, gold_answers), which returns the max EM and the max F1 over all gold answers: {"em": ..., "f1": ...}.