Add LLM compose stack for phy-srv-gpu01 (not deployed yet)

Follows the homelab pattern: ironicbadger.docker_compose_generator v2
renders services/<host>/NN-<stack>/compose.yml templates into
~/docker/compose.yaml on the host.

- 01-vllm: chat model, fixed --gpu-memory-utilization
- 02-embeddings: second vLLM instance (--task embed) rather than a
  separate toolchain, so SM120 support only has to be solved once
- 03-openwebui: Open WebUI + pgvector (not chroma — corpus size)
- 99-network: shared bridge; leading comment keeps networks: top-level
- pin docker_compose_generator to 2.0.1 — galaxy tags mix v1/v2 formats
- group_vars: stack config incl. LDAP placeholders still to be filled

The role only writes the compose file; starting the stack stays manual.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

ansible/services/phy-srv-gpu01/02-embeddings/compose.yml

@@ -0,0 +1,37 @@
+services:
+  # Embeddings run on a second vLLM instance rather than a separate toolchain
+  # (e.g. text-embeddings-inference): whatever vLLM build works on SM120 then
+  # covers embeddings too, instead of having to solve Blackwell support twice.
+  embeddings:
+    image: "{{ vllm_image }}"
+    container_name: embeddings
+    networks:
+      - llmnet
+    ports:
+      - "8001:8000"
+    volumes:
+      - "{{ appdata_path }}/models/huggingface:/root/.cache/huggingface"
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all
+      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
+      - "HUGGING_FACE_HUB_TOKEN={{ hf_token | default('') }}"
+    command:
+      - --model
+      - "{{ embedding_model }}"
+      - --served-model-name
+      - "{{ embedding_model_name }}"
+      - --task
+      - embed
+      # small fixed slice — the chat model gets the rest
+      - --gpu-memory-utilization
+      - "{{ embeddings_gpu_memory_utilization }}"
+    ipc: host
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    runtime: nvidia
+    restart: unless-stopped