feat(US5+6): QA generation — POST /api/v1/qa/gen-text and /gen-image

- Add qa_models.py with TextQAItem, GenTextQARequest, QAPair, ImageQAItem,
  GenImageQARequest, ImageQAPair, TextQAResponse, ImageQAResponse
- Implement gen_text_qa(): batch-formats triples into a single prompt, calls
  llm.chat(), parses JSON array via extract_json()
- Implement gen_image_qa(): downloads cropped image from source-data bucket,
  base64-encodes inline (data URI), builds multimodal message, calls
  llm.chat_vision(), parses JSON; image_path preserved on ImageQAPair
- Replace qa.py stub with full router: POST /qa/gen-text and /qa/gen-image
  using Depends(get_llm_client) and Depends(get_storage_client)
- 15 new tests (8 service + 7 router), 53/53 total passing

app/models/qa_models.py

@@ -0,0 +1,47 @@
+from pydantic import BaseModel
+
+
+class TextQAItem(BaseModel):
+    subject: str
+    predicate: str
+    object: str
+    source_snippet: str
+
+
+class GenTextQARequest(BaseModel):
+    items: list[TextQAItem]
+    model: str | None = None
+    prompt_template: str | None = None
+
+
+class QAPair(BaseModel):
+    question: str
+    answer: str
+
+
+class ImageQAItem(BaseModel):
+    subject: str
+    predicate: str
+    object: str
+    qualifier: str | None = None
+    cropped_image_path: str
+
+
+class GenImageQARequest(BaseModel):
+    items: list[ImageQAItem]
+    model: str | None = None
+    prompt_template: str | None = None
+
+
+class ImageQAPair(BaseModel):
+    question: str
+    answer: str
+    image_path: str
+
+
+class TextQAResponse(BaseModel):
+    pairs: list[QAPair]
+
+
+class ImageQAResponse(BaseModel):
+    pairs: list[ImageQAPair]