fix distributed evaluation on empty task shards

[Luodian]

Summary

• Fix distributed evaluation when a rank receives zero documents after task sharding.
• Keep request construction and synchronization aligned by injecting a padding request for empty shards.
• Scope the change to task request building and evaluator synchronization.

In scope

• Update lmms_eval/api/task.py to preserve distributed synchronization for empty shards.
• Update lmms_eval/evaluator.py to keep request/filter coordination consistent across ranks.
• Include the formatter adjustments required by the existing lint rules.

Out of scope

• vLLM backend dispatch changes.
• Benchmark-specific scoring or prompt changes.

Validation

• uv run --with pytest python -m pytest test/eval/test_construct_requests.py -q | sample size: N=23 tests | key metrics: 23 passed | result: pass
• uv run python - <<'PY' import lmms_eval.evaluator as evaluator print(hasattr(evaluator, 'evaluate')) PY | sample size: N=1 smoke check | key metrics: evaluate=True | result: pass
• uv run pre-commit run --all-files | sample size: N=all tracked files | key metrics: black, isort passed | result: pass

Risk / Compatibility

• Moderate behavior change: distributed runs with empty shards now stay alive instead of diverging or hanging.
• The new padding path only applies when a rank has no documents, so normal single-rank and balanced runs are unaffected.

Type of Change

• Bug fix (non-breaking change)
• New feature
• New benchmark/task
• New model integration
• Breaking change
• Documentation update
• Refactoring (no functional changes)