Add CyberGym benchmark integration

[bmz-q-q]

What

• add a new qitos.benchmark.cybergym benchmark family
• add qitos.recipes.benchmarks.cybergym and a thin examples/benchmarks/cybergym_eval.py entrypoint
• vendor the current cybergym_agent under the CyberGym benchmark package
• add Chinese benchmark docs and register CyberGym in the Chinese benchmark overview
• add unit coverage for task loading, trace writer setup, benchmark registration, GLM family mapping, and submit-tool normalization

Verification

• python -m unittest tests.test_benchmark_cybergym_recipe
• smoke run with examples/benchmarks/cybergym_eval.py on arvo:1065, writing trace artifacts under runs/cybergym/traces

Notes

• public docs have been scrubbed of concrete private IPs, keys, and endpoint domains
• current smoke reaches the runner and writes QitOS traces, but GLM-5.1-sii still emits a tool-call format that does not yet align with the current JsonDecisionParser; that protocol alignment is a follow-up task rather than a benchmark-structure blocker