Graphiti's OpenAIClient routes "simpler" extraction prompts to
ModelSize.small, falling back to a hard-coded `gpt-4.1-nano` when no
small_model is configured. Against Qwen/Dashscope (and other OpenAI-SDK
endpoints that don't host that model) this surfaces as a 404 mid-graph
build, even when LLM_MODEL_NAME points at a working model.
Add LLM_SMALL_MODEL_NAME (defaults to LLM_MODEL_NAME) and pass it through
the Graphiti LLMConfig so the small slot is always populated with a
model the configured endpoint actually serves.
Closes#37
8af800a122