Capture SWE-bench patches on failed runs

[neubig]

Summary

• add a generic failure-result hook that runs before workspace cleanup
• use the hook in SWE-bench to stage workspace changes and recover a git diff from failed/stuck runs
• preserve recovered patches in error rows via test_result.git_patch so they can still be scored

Validation

• /home/gneubig/work/openhands-benchmarks-venv/bin/python -m pytest -q tests/test_swebench_failure_patch_capture.py tests/test_error_output_serialization.py
• /home/gneubig/work/openhands-benchmarks-venv/bin/python -m ruff check benchmarks/utils/evaluation.py benchmarks/swebench/run_infer.py tests/test_swebench_failure_patch_capture.py
• /home/gneubig/work/openhands-benchmarks-venv/bin/python -m ruff format --check benchmarks/utils/evaluation.py benchmarks/swebench/run_infer.py tests/test_swebench_failure_patch_capture.py
• /home/gneubig/work/openhands-benchmarks-venv/bin/python -m py_compile benchmarks/utils/evaluation.py benchmarks/swebench/run_infer.py tests/test_swebench_failure_patch_capture.py

This PR was created by Codex on behalf of Graham Neubig.

Issue

Closes #749