Add multi-provider support via LiteLLM, removing hard dependency on O…

[getmykhan]

Add multi-provider support via LiteLLM, removing hard dependency on OpenRouter

Previously, all LLM calls were routed exclusively through OpenRouter using
raw httpx requests. This change introduces LiteLLM as a unified LLM client,
allowing users to call provider APIs directly (OpenAI, Anthropic, Google
Gemini, xAI, DeepSeek, etc.) by setting the corresponding API key in .env.
OpenRouter is preserved as an automatic fallback for any model that lacks a
direct provider key.

Why this matters:

• Lower latency: direct API calls skip the OpenRouter proxy
• Lower cost: no OpenRouter markup on per-token pricing
• Flexibility: users can mix and match — e.g. use OpenAI directly but
  route xAI through OpenRouter
• No breaking change: setting only OPENROUTER_API_KEY still works exactly
  as before

How routing works:

1. resolve_model() checks if a direct API key exists for the model's
   provider (e.g. OPENAI_API_KEY for openai/* models)
2. If yes, the model is passed to litellm as-is (native API call)
3. If no, the model is prefixed with "openrouter/" and the provider
   prefix is translated (e.g. gemini/ -> google/, xai/ -> x-ai/) to
   match OpenRouter's naming convention
4. If neither a direct key nor OPENROUTER_API_KEY is set, a clear
   ValueError is raised telling the user which env var to set

Changes:

• backend/llm_client.py: New unified LLM client using litellm.acompletion(),
  replacing the httpx-based openrouter.py. Same query_model() and
  query_models_parallel() interface for zero-touch migration.
• backend/config.py: Added PROVIDER_KEY_MAP, LITELLM_TO_OPENROUTER_PROVIDER,
  resolve_model(), and print_routing_info(). Model identifiers now use
  litellm naming convention (e.g. "gemini/" instead of "google/"). Added
  UTILITY_MODEL config for lightweight tasks (title generation).
• backend/council.py: Updated imports from openrouter -> llm_client. Title
  generation now uses configurable UTILITY_MODEL instead of hardcoded
  OpenRouter model name.
• backend/main.py: Migrated from deprecated @app.on_event("startup") to
  FastAPI lifespan context manager. Prints provider routing table on startup.
• pyproject.toml: Added litellm>=1.60.0 dependency.
• .env.example: New file documenting all supported provider API keys with
  usage instructions.
• CLAUDE.md: Updated architecture docs to reflect multi-provider support.
• backend/openrouter.py: Kept for reference but no longer imported.