fix(graphrag-llm): accept non-OpenAI service_tier values in LLMCompletionResponse

[LijuanTang94]

Closes #2389

Problem

Non-OpenAI providers routed through LiteLLM (e.g. Gemini via OpenRouter) return a service_tier value outside OpenAI's strict literal (auto/default/flex/scale/priority). Because LLMCompletionResponse extends openai.types.chat.ChatCompletion, constructing it via LLMCompletionResponse(**response.model_dump()) in lite_llm_completion.py fails with a pydantic literal_error, breaking the extract_graph workflow:

1 validation error for LLMCompletionResponse
service_tier
  Input should be 'auto', 'default', 'flex', 'scale' or 'priority'

Fix

service_tier is an OpenAI-specific field that graphrag-llm does not use. This overrides it in the LLMCompletionResponse subclass to accept str | None, so non-standard provider values no longer fail validation. This is more robust than stripping the field on a single conversion path, as it covers every place the model is constructed.

Verification

• Constructing LLMCompletionResponse with a non-OpenAI service_tier (e.g. an OpenRouter value) now succeeds instead of raising.
• Standard OpenAI values (default) and None still work.
• The content computed field is unaffected.
• ruff check / ruff format --check pass; semversioner change file added.

[LijuanTang94]

@microsoft-github-policy-service agree