feat: add Google Gemini provider support to memU

[loki52501]

📝 Pull Request Summary

Add Google Gemini as a supported LLM and embedding provider via its OpenAI-compatible API endpoint.

✅ What does this PR do?

• Adds GeminiLLMBackend in llm/backends/gemini.py that inherits from OpenAILLMBackend, using Gemini's
  /v1beta/openai/ OpenAI-compatible endpoint
• Registers gemini in LLM_BACKENDS and the embedding backends map in llm/http_client.py
• Adds provider="gemini" auto-defaults in LLMConfig.set_provider_defaults() — sets correct base_url,
  chat_model (gemini-2.5-flash), and embed_model (gemini-embedding-001) automatically
• Adds a full test suite in tests/llm/test_gemini_provider.py covering settings defaults, backend
  registration, payload building, HTTP client initialization, and live API calls
• Updates README.md with a Google Gemini Integration section

---

🤔 Why is this change needed?

• Gemini is a widely used LLM provider with a free tier, making it a practical default for self-hosted
  memU deployments
• Users running memU locally without an OpenAI subscription have no built-in alternative — this gives
  them a zero-cost path to get started
• Gemini's /v1beta/openai/ endpoint is fully OpenAI-compatible, so no new HTTP client or request format
  was needed — the change is minimal and low-risk

---

🔍 Type of Change

• Bug fix
• New feature
• Documentation update
• Refactor / cleanup
• Other

---

✅ PR Quality Checklist

• PR title follows the conventional format (feat: add Google Gemini provider support)
• Changes are limited in scope and easy to review (4 files modified, 2 files added)
• Documentation updated — Gemini section added to README.md matching OpenRouter format
• No breaking changes — existing providers (openai, grok, openrouter, doubao) are unaffected
• Related issues or discussions linked — N/A (standalone addition)

---

📌 Optional

• Live API tests confirmed passing — test_chat, test_summarize, test_embed all return 200 OK with
  gemini-2.5-flash and gemini-embedding-001
• Edge cases considered — explicit config values are not overridden by defaults; free-tier quota limits
  noted in docs (gemini-2.0-flash has lower RPD than gemini-2.5-flash)
• Follow-up tasks: consider adding gemini-embedding-exp-03-07 to the supported models table once it
  leaves experimental status