fix: compose env benchmark owner config

[TATP-233]

Summary

• compose every env-step benchmark task from its Hydra owner configuration
• remove the drifted hand-written Go2W rough reward mirror
• pass env.* and reward.* overrides through Hydra and reject standalone training.sim_backend
• add config-boundary regression coverage for owner parity, interpolation, typing, and invalid overrides

Scope

• Parent umbrella: Work: 代码清理审计 umbrella：边界契约/结构契约/死代码/重复实现/腐化代码结论汇总(只审计) #983
• Resolves child issue Fix: 让 env-step benchmark 通过 Hydra owner 配置构造任务 #999 after this PR is merged into dev/issue-983-code-cleanup
• Integration target: dev/issue-983-code-cleanup only; no change targets main

Validation

• focused tests: 16 passed
• all 27 declared task/backend config combinations composed and validated
• PYTHONPATH="$PWD/src" UV_NO_SYNC=1 make test-all
• pytest: 1681 passed, 27 skipped, 273 deselected, 1 xfailed
• benchmark smoke: 33/34 module imports and 34/35 script imports; only platform-optional MLX skipped
• Ruff, mypy, and pyright passed
• exact validated head: 9f7cff39fac275eb65ca11f6994d7fab659e5ebe