Fix request-level workspace routing

[Pioneer-Weirdo]

Summary

• route query, document, graph, and Ollama API requests through a workspace-aware LightRAG registry
• add /workspaces inventory for runtime inspection and make /health initialize/report the active workspace consistently
• avoid mutating a shared rag.workspace in-place by lazily creating one LightRAG instance per workspace

Problem

LIGHTRAG-WORKSPACE was parsed by the server, but only /health used it. Query and write routes still operated on the single startup rag instance, so request-level workspace isolation did not actually work.

Changes

• add get_rag_for_workspace(...) and per-workspace DocumentManager resolution in lightrag_server.py
• pass workspace resolvers into query/document/graph routers and OllamaAPI
• update handlers to accept Request and resolve the active workspace-specific LightRAG before serving the request
• finalize all initialized workspace instances on shutdown
• expose /workspaces to inspect known workspaces discovered from config, registries, and storage directories

Verification

• python -m py_compile lightrag/api/lightrag_server.py lightrag/api/routers/query_routes.py lightrag/api/routers/document_routes.py lightrag/api/routers/graph_routes.py lightrag/api/routers/ollama_api.py
• runtime A/B workspace probe against a patched LightRAG container confirmed:
  • writing token A to workspace A is only visible in workspace A
  • writing token B to workspace B is only visible in workspace B

[github-actions]

This pull request has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.

[github-actions]

This pull request has been automatically closed because it has not had recent activity.