feat(llm): extract BaseAnthropicLLM, add base_url, export extension points

[matteomedioli]

Description

Stacked on top of #565 (the http_client sync/async fix), this PR extends the same "shared base class, thin provider subclass" shape already used for OpenAI (BaseOpenAILLM → OpenAILLM/AzureOpenAILLM) to Anthropic.

AnthropicLLM today mixes two concerns in one class: provider-shape logic (message building, JSON-schema conversion, response parsing) and SDK client construction. There's no supported way to reuse the provider-shape logic against a different Anthropic-compatible endpoint without copying the whole class.

Changes:

• Extract BaseAnthropicLLM(LLMBase, abc.ABC) holding all of AnthropicLLM's message-building, schema-conversion, and response-parsing logic, moved verbatim — no behavior change.
• AnthropicLLM becomes a thin subclass of BaseAnthropicLLM whose __init__ only constructs anthropic.Anthropic/anthropic.AsyncAnthropic (reusing the sync/async http_client split from fix(llm): route http_client to the matching sync/async Anthropic client #565).
• Add an explicit, documented base_url: str | None = None keyword parameter to AnthropicLLM, passed through to both SDK clients — previously only reachable as an undocumented **kwargs passthrough accident.
• Export both BaseAnthropicLLM and BaseOpenAILLM from neo4j_graphrag.llm and add them to __all__ — BaseOpenAILLM already existed but was never exported, so this also fixes that gap for OpenAI as a side effect of applying one consistent policy to both families.
• Add a new docs page (docs/source/llm_extensibility.rst, linked from the docs index and API reference) documenting the http_client/base_url contract shared by both base classes, with a worked subclassing example.

Every provider SDK already supports pointing at a custom endpoint over a custom transport — the limitation was never at the SDK level, only in how these classes exposed (or didn't expose) that as a documented constructor contract. With this change, anyone running a model behind an Anthropic-compatible endpoint — a self-hosted gateway, an internal proxy, a compatibility layer, or simply a different region/deployment — can write a small subclass instead of forking AnthropicLLM or relying on **kwargs accidents:

from neo4j_graphrag.llm import BaseAnthropicLLM
import anthropic

class MyCustomAnthropicLLM(BaseAnthropicLLM):
    def __init__(self, model_name, **kwargs):
        super().__init__(model_name=model_name, **kwargs)
        self.client = anthropic.Anthropic(
            base_url="https://my-endpoint.example.com",
            http_client=my_configured_httpx_client,
        )
        self.async_client = anthropic.AsyncAnthropic(
            base_url="https://my-endpoint.example.com",
            http_client=my_configured_async_httpx_client,
        )

All message-building, schema conversion, and response parsing is inherited — nothing provider-shape-related needs reimplementing.

Only 2 of our 8 LLM providers follow this shape today — AnthropicLLM/BaseAnthropicLLM (this PR) and OpenAILLM/AzureOpenAILLM/BaseOpenAILLM (pre-existing). The other six — CohereLLM, MistralAILLM, OllamaLLM, BedrockLLM, GeminiLLM, and VertexAILLM — still mix provider-shape logic and SDK client construction in one class, with no Base<X>LLM split and no exported extension point. Worth a follow-up pass to align each of them where the provider's SDK actually has a client object to factor out; a couple, like Bedrock (boto3 session/endpoint_url, not an httpx client) and VertexAI (no per-instance client at all, driven by global vertexai.init()), will need a different-shaped extension point. Not proposing that fix here, just flagging the inconsistency as worth scoping separately.

Type of Change

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

Complexity

Complexity: Medium

How Has This Been Tested?

• Unit tests
• E2E tests
• Manual tests

BaseAnthropicLLM holds all previously-AnthropicLLM logic, unchanged in behavior; base_url verified passed through to both anthropic.Anthropic and anthropic.AsyncAnthropic; BaseAnthropicLLM/BaseOpenAILLM verified importable from neo4j_graphrag.llm and present in __all__; full tests/unit/llm/ suite green; mypy/ruff clean; new docs page builds without new Sphinx warnings.

Checklist

The following requirements should have been met (depending on the changes in the branch):

• Documentation has been updated
• Unit tests have been updated
• E2E tests have been updated
• Examples have been updated
• New files have copyright header
• CLA (https://neo4j.com/developer/cla/) has been signed
• CHANGELOG.md updated if appropriate