feat(llm): extract BaseGeminiLLM, add base_url support

[matteomedioli]

Stack

Merge order: #565 → #567 → #566 → #571 / #572 (last two in either order). Each PR builds on the previous one. Until its base merges, GitHub may show parent commits in the diff — review only the commits unique to this branch.

Description

Same split as #566, applied to Gemini.

• New BaseGeminiLLM holds all the shared logic: message building, config and schema building, response parsing, tool handling. GeminiLLM becomes a thin subclass whose only job is creating the genai.Client.
• BaseGeminiLLM is exported from neo4j_graphrag.llm as a supported extension point. A subclass reaching a custom endpoint just builds its own client — everything else is inherited. A test proves the contract: a minimal subclass that only assigns a client runs invoke() end to end.
• New explicit base_url parameter on GeminiLLM, mirroring feat(llm): extract BaseAnthropicLLM, add base_url, export extension points #566. The google-genai SDK has no top-level base_url argument, so the value is applied through http_options; when http_options is also passed (dict or types.HttpOptions), only its base_url field is overridden. All three merge paths are tested, plus the negative path (no base_url → nothing passed).
• aclose() now closes the genai.Client (sync and aio surfaces), so with GeminiLLM(...) / async with actually release resources like the other providers.
• Docs: llm.rst extensibility page extended with a Gemini section (single-client model, http_options instead of http_client), api.rst cross-references, base_url example added.

Behavior notes:

• The extraction itself is behavior-neutral (commits 1–2); base_url and aclose are small additive features on top.
• An earlier version of this PR also enabled the supports_structured_output flag for Gemini. That silently changes pipeline defaults for existing users, so it moved to its own PR (feat(llm): GeminiLLM declares supports_structured_output = True #573) with a proper breaking-change note.
• GeminiLLM's parent class changes from the two interface classes to LLMBase, which extends both — every existing isinstance check still passes, no deprecation warning fires, and the class gains working context-manager support (with GeminiLLM(...) as llm:) like the other providers.

Type of Change

• New feature
• Bug fix
• Breaking change
• Documentation update
• Project configuration change

Complexity

Complexity: Low

How Has This Been Tested?

• Unit tests
• E2E tests
• Manual tests

Checklist

• Documentation has been updated
• Unit tests have been updated
• Examples have been updated
• CLA (https://neo4j.com/developer/cla/) has been signed
• CHANGELOG.md updated

[stellasia]

Weird that we need both. Do they also have both a sync and an async connection internally?

[matteomedioli]

Seems so:

• Client.close() and AsyncClient.aclose()
  https://github.com/googleapis/python-genai/blob/v1.62.0/google/genai/client.py

• BaseApiClient.close() / aclose():
  https://github.com/googleapis/python-genai/blob/v1.62.0/google/genai/_api_client.py

I run a quick minimal test:

import asyncio
from google import genai

async def main() -> None:
    client = genai.Client(api_key="fake")
    api = client._api_client
    client.close()
    print("after close():       sync closed:", api._httpx_client.is_closed,
          "| async closed:", api._async_httpx_client.is_closed)  # True | False
    await client.aio.aclose()
    print("after aio.aclose():  sync closed:", api._httpx_client.is_closed,
          "| async closed:", api._async_httpx_client.is_closed)  # True | True
asyncio.run(main())