Docker Compose: Fix placement constraints for non-GPU OPEA wrappers

Parent: #758 (closed)

Summary

Move placement constraints for non-GPU services off the GPU node. Only the actual model servers that load weights into VRAM should require GPU placement.

Current Problem

These services have gpu == true placement but don't need GPU:

• embedding: Thin proxy → calls TEI via TEI_EMBEDDING_ENDPOINT (HTTP)
• reranker: Thin proxy → calls TEI via TEI_RERANKING_ENDPOINT (HTTP)
• translation: Thin proxy → calls vLLM via LLM_ENDPOINT (HTTP)
• textgen: Passthrough → calls vLLM via VLLM_ENDPOINT (HTTP)

Additionally, these services are on the GPU node but don't need GPU:

• dataprep-arango-service: Calls vLLM, TEI, and (with #764 (closed)) docling-service all via HTTP. Only the in-process docling needs GPU, which is being extracted to its own service.
• retriever-arango-service: Calls vLLM and TEI via HTTP endpoints
• chatqna-xeon-backend-server: Calls all model services via HTTP endpoints

Services that KEEP gpu == true

Only services that load models into GPU memory:

• vllm — loads LLM into VRAM
• vllm-translation-guardrail — loads translation model into VRAM
• tei — loads embedding model into VRAM
• tei_reranker — loads reranker model into VRAM
• docling-service (NEW, from #764 (closed)) — loads docling + easyocr models into VRAM

Services that MOVE to genieai == true

embedding:
  deploy:
    placement:
      constraints:
        - node.labels.genieai == true

reranker:
  deploy:
    placement:
      constraints:
        - node.labels.genieai == true

translation:
  deploy:
    placement:
      constraints:
        - node.labels.genieai == true

textgen:
  deploy:
    placement:
      constraints:
        - node.labels.genieai == true

dataprep-arango-service:
  deploy:
    placement:
      constraints:
        - node.labels.genieai == true

retriever-arango-service:
  deploy:
    placement:
      constraints:
        - node.labels.genieai == true

chatqna-xeon-backend-server:
  deploy:
    placement:
      constraints:
        - node.labels.genieai == true

Files

File	Change
docker-compose.yaml	Update placement constraints for 7 services

Acceptance Criteria

• 5 model server services keep gpu == true placement
• 7 non-GPU services use genieai == true placement
• Single-node deployment (both labels on same node) works unchanged
• App node runs all 7 non-GPU services, calling remote GPU model endpoints