Fix Metrics.drop corpus aggregation to use mean instead of max

[salignatmoandal]

Summary

• Change Metrics.drop corpus aggregation from max to np.mean for both em and f1.
• Keep per-sample max over alternate gold answers inside DropMetrics; only the across-document aggregation was wrong.
• Add a regression test so one correct document among four reports 0.25, not 1.0.

Closes #1395.

Context

DropMetrics.compute already takes the max over gold spans for a single document. MetricsLogger.aggregate then applied corpus_level_fn={"em": max, "f1": max}, so any single exact hit made the reported task score 1.0. That diverges from lm-eval-harness DROP (aggregation: mean) and from Metrics.exact_match in the same file.

The built-in lighteval|drop task uses Metrics.exact_match, so its numbers are unaffected. Custom tasks that select Metrics.drop were reporting inflated scores.

Test plan

• pytest tests/unit/metrics/test_drop_aggregation.py — 2 passed
• ruff check / ruff format --check on changed files
• CI unit tests on the PR