Optimize tau-bench driver hot paths

[bradhilton]

Summary

• raise tau environment and policy HTTP connection capacity for highly concurrent rollouts
• avoid redundant trajectory/completion model dumps during history reconciliation and tensorization
• reduce recursive pre-pickle string interning overhead while preserving object identity and cycle handling

Companion service/driver changes: OpenPipe/caladan#144.

Testing

• pytest -q tests/unit/test_tau_bench_client.py tests/unit/trajectories/test_history.py tests/unit/trajectories/test_tensorized_models.py tests/unit/trajectories/test_tokenized_models.py (91 passed, 1 skipped)
• changed-file Ruff check and format check
• changed-file ty check
• GitHub quality checks and install smoke test