fix: harden AutoRL-Bench RL evaluators

[couragec]

Summary

This PR hardens the AutoRL-Bench RL evaluators by backporting a small set
of targeted fixes:

• Fix ALFWorld vLLM evaluation crashes from overlong ReAct history by
  truncating prompts to fit the configured context window.
• Ensure ALFWorld restores stdout and cleans up the environment / vLLM
  backend even when evaluation fails.
• Propagate vLLM-safe environment variables to OpenCompass subprocesses
  and set enforce_eager=True in the OpenCompass template.
• Prevent failed baseline evaluations from silently becoming cached or
  returned as valid 0.0 scores.
• Improve timeout cleanup by killing the process session/tree, while
  keeping kill_process_group as a compatibility alias.

Validation

• python -m py_compile rdagent/scenarios/rl/autorl_bench/benchmarks/ alfworld/eval.py rdagent/scenarios/rl/autorl_bench/core/evaluator.py rdagent/scenarios/rl/autorl_bench/core/utils.py rdagent/scenarios/rl/ autorl_bench/core/opencompass.py rdagent/scenarios/rl/autorl_bench/core/ __init__.py rdagent/scenarios/rl/autorl_bench/test/test_fixes.py
• ruff check --no-fix --select F,E9 rdagent/scenarios/rl/autorl_bench/ benchmarks/alfworld/eval.py rdagent/scenarios/rl/autorl_bench/core/ evaluator.py rdagent/scenarios/rl/autorl_bench/core/utils.py rdagent/ scenarios/rl/autorl_bench/core/opencompass.py rdagent/scenarios/rl/ autorl_bench/core/__init__.py rdagent/scenarios/rl/autorl_bench/test/ test_fixes.py
• /tmp/rdagent-test-venv/bin/python -m rdagent.scenarios.rl.autorl_bench.test.test_fixes

Result: 24 passed, 0 failed.

---

📚 Documentation preview 📚: https://RDAgent--1402.org.readthedocs.build/en/1402/