Support qwen_chat as optimizer backend

[LifeIsSoSolong]

Summary

• allow qwen_chat to be used as an optimizer backend, not only as a target backend
• add role-specific Qwen/vLLM configuration for optimizer and target endpoints
• keep existing shared QWEN_CHAT_* behavior as the default fallback
• update the Qwen target OpenAI client path to use the new target config, which keeps direct OpenAI-client benchmark paths compatible

Details

This makes it possible to run local OpenAI-compatible vLLM services for both the SkillOpt optimizer and the target model, including setups where optimizer and target are served by different Qwen models or different endpoints.

Existing Qwen target usage remains compatible: QWEN_CHAT_BASE_URL, QWEN_CHAT_API_KEY, QWEN_CHAT_MODEL, and other shared QWEN_CHAT_* settings continue to act as shared defaults. Role-specific OPTIMIZER_QWEN_CHAT_* and TARGET_QWEN_CHAT_* values override those defaults when provided.

Validation

• git diff --check
• python -m py_compile scripts/train.py skillopt/config.py skillopt/engine/trainer.py skillopt/model/__init__.py skillopt/model/backend_config.py skillopt/model/qwen_backend.py skillopt/model/azure_openai.py
• python scripts/train.py --help
• local config assertions for shared QWEN_CHAT_MODEL and role-specific Qwen endpoint/thinking settings

[LifeIsSoSolong]

@microsoft-github-policy-service agree