feat: add multimodal embedding service for IBM Power (ppc64le, CPU-only)

[nava-dba]

💡 Feature Description

As a platform operator running vector-capable databases, I want a
native embedding service in the ai-services stack so that I can
generate vectors locally on IBM Power without depending on a remote
endpoint or managing a custom container outside the catalog.

Currently there is no packaged embedding generation service in the
ai-services stack. Teams using vector databases must operate a
separate embedding service themselves or call a remote endpoint.
This feature adds embedding generation as a first-class ai-service,
consistent with how other services (summarize, similarity, chatbot)
are deployed and managed today.

The service exposes an OpenAI-compatible POST /v1/embeddings
endpoint. It runs CPU-only on IBM Power11 ppc64le (RHEL 9, no Spyre
cards required), deployable via the ai-services catalog UI using
Podman or as a standalone Linux sidecar container alongside any
application stack. Any vector database framework that consumes an
OpenAI-compatible embeddings endpoint can use this service.

❗ Why This Matters

Teams using vector databases (OpenSearch, pgvector, Oracle, Milvus,
etc.) need an embedding service to generate vectors from text or
images. Without a native option in the stack, every team must:

• Call a remote endpoint (latency, data leaves the system), or
• Manage a custom container outside the ai-services catalog

Embedding generation is a core primitive for any RAG or vector search
workflow. Adding it as a native service means:

• Any vector database on Power can use it through a standard API
• No Spyre cards required — works on any Power11 LPAR
• Consistent deployment model with the rest of the ai-services catalog
• Can be offered as a standalone service or bundled in architectures

🧩 Requirements

• Expose POST /v1/embeddings compatible with the OpenAI embeddings
  API (text and image inputs)
• Support models validated by DMG and compatible with the vLLM runtime
  (e.g. clip-vit-base-patch32, siglip-base-patch16-224)
• Model selection via MODEL_NAME environment variable
• Run CPU-only on IBM Power11 ppc64le RHEL 9 — no Spyre cards required
• Deployable via Podman through the ai-services catalog UI
• Return float32 vectors consumable by any vector database framework
• Include unit tests with mocked model inference (no hardware required
  to run CI)
• Pass existing CI workflows: go.yaml and service-unit-tests.yml

📚 Additional Details

• Jira: AISERVICES-1811
• Reference implementation tested on IBM Power11 LPAR, RHEL 9.6,
  Podman runtime
• Fork with working implementation:
  https://github.com/nava-dba/embedding-services
• Release binary tested: v0.4.11-embedding (ai-services-linux-ppc64le)
• No new global infrastructure required — service is self-contained,
  no OpenSearch or PostgreSQL dependency