Fix Dyck scoring to preserve bracket order

[Excelius-Wang]

Summary

• replace inspect-ai's punctuation-normalizing exact() scorer for Dyck tasks with a task-local
  whitespace-trimmed comparison
• preserve bracket order and multiplicity for the two-, three-, and four-bracket variants
• add regression coverage for exact output, surrounding whitespace, reordered brackets, and
  duplicated brackets

Fixes #1329.

Why

Raising the inspect-ai dependency floor does not fully fix bracket-only answers: inspect-ai 0.3.260
still normalizes reordered bracket sequences to the same value. The task-local scorer mirrors
lighteval's existing exact-match behavior by trimming surrounding whitespace only, while keeping the
existing mean and stderr metrics.

Testing

• pytest -q tests/unit/tasks/test_dyck_language.py
• pytest -q tests/unit/tasks/test_lighteval_task.py::test_hf_data_files tests/unit/tasks/templates tests/unit/tasks/test_dyck_language.py
• ruff check src/lighteval/tasks/tasks/dyck_language.py tests/unit/tasks/test_dyck_language.py
• ruff format --check src/lighteval/tasks/tasks/dyck_language.py tests/unit/tasks/test_dyck_language.py
• git diff --check