feat(tests): add TraceLab coding-agent benchmark

[GentleCold]

Summary

• Export TraceLab coding-agent sessions into the official closed-loop CSV shape while preserving token lengths, round order, and tool gaps.
• Generate model-native token pools for exact token-ID replay across rounds.
• Document the external dataset/runner workflow and add focused selection and serialization tests.

Motivation

Long-context agent workloads need session-level timing and prefix growth rather than independent synthetic prompts. This adapter makes the TraceLab workload reproducible without vendoring its dataset or runner.

Test plan

• Unit tests added / updated (5 passed focused)
• PYTHONHASHSEED=0 pytest -q -m "not integration" --ignore=tests/integration --ignore=tests/connector/test_daser_connector.py --ignore=tests/connector/test_gds_transfer.py tests/ passes (358 passed)
• Integration tests marked N/A; the dataset and runner remain external
• Real DuckDB smoke: 2 sessions, 25 rounds, model-derived capacity summary emitted
• GLM tokenizer smoke: 28 valid model-native token IDs serialized and read back
• pre-commit run --all-files passes

Checklist

• pre-commit run --all-files passes
• Type hints and docstrings on all new/modified functions
• New behavior includes tests
• Existing CPU test suite passes