feat(runner): add optional timing collection for grading workflows

[Alexxigang]

Summary

• add a reusable TimingCollector utility under openjudge/utils/timer.py with context-manager and decorator support
• instrument GradingRunner to optionally collect latency records for single evaluations, whole datasets, and multi-dataset runs
• expose timing records and summaries through get_timing_records(), get_timing_summary(), and clear_timing_records()
• add unit tests for the timing utility and runner timing integration

Why this fix

OpenJudge did not have a built-in way to measure the latency of core evaluation workflow steps, which made it harder to identify bottlenecks or track regressions over time.

This PR introduces a lightweight timing utility that logs at DEBUG level by default and stores in-memory timing records for programmatic access. As a pilot integration, the grading workflow now supports optional timing collection without changing existing behavior when timing is disabled.

Closes #81.

Validation

• PYTHONPATH=. pytest tests/utils/test_timer.py tests/runner/test_grading_runner.py -q