feat(swe): SWE-bench evaluation with AGS sandbox

[Lightblues]

Summary

• Add SWERexEnv for remote coding in AGS sandboxes with SWE-ReX runtime (bash, read_file, write_file, edit_file tools)
• Implement SWEBenchmark with per-instance Docker image lifecycle and patch extraction
• Add data import/export scripts and evaluation config for SWE-bench Verified/Lite/Full

Changes

• utu/env/swerex_env.py: AGS sandbox environment with pager-disable fix
• utu/eval/benchmarks/swe_benchmark.py: Per-instance rollout with deep-copy config, git diff via file API
• utu/eval/processer/swe_bench.py: Preprocessor with prompt template, patch_rate metrics
• scripts/data/process_swe_bench.py: HuggingFace → DB import with image name mapping
• scripts/data/export_swe_bench_patches.py: DB → JSONL export for offline evaluation
• configs/: env, agent, eval YAML configs
• tests/env/test_swerex_env.py: 25 unit tests with mocked runtime
• examples/swe/swe-bench.md: Usage documentation

Closes #251