fix(ai): route scene generation through configured providers

[fj-sunny]

Summary

• route custom scene generation through the admin-configured LLM route
• prioritize Qiniu MaaS and allow fallback to Alibaba Qwen when Qiniu rejects a request
• route model-less TTS calls through the configured TTS fallback chain
• add the admin fallback-model control and remove the Qiniu environment kill switch
• use dynamic Alibaba workspace credentials and assign distinct audit request IDs

Tests

• ./mvnw --settings docker/maven/settings.xml -Dtest=AiProviderRegistryTest,CustomSceneGeneratorTest,CustomSceneServiceTest,QwenLlmProviderTest test (59 passed)
• npm test -- --run src/features/system/SystemManagementPage.test.tsx (8 passed)
• npm run build

Closes #182

[fennoai]

Review conclusion

I reviewed the complete fixed diff for the configurable LLM/TTS routing, Qiniu-to-Qwen failover, dynamic Alibaba workspace endpoint handling, audit request scoping, and admin fallback-model controls. I did not identify a reportable correctness or regression issue in the changed code. git diff --check passes. The focused Maven tests could not start because the configured ${env.MAVEN_REPO_URL} repository is unavailable in this environment, and the admin test dependencies are not installed (node_modules is absent).

[pionxe]

LGTM