Persist LocalAI state and match the live deployment

- Keep agents, KB collections, users and runtime settings
  (LOCALAI_DATA_PATH, LOCALAI_CONFIG_DIR) and backends
  (LOCALAI_BACKENDS_PATH) on the models PVC. The chart's /data,
  /configuration and /backends are ephemeral, so every pod restart wiped
  the agent team and the llama.cpp backend.
- Default the agent pool to ornith-1.5-9b-q8 with
  granite-embedding-107m-multilingual embeddings.
- Make the image tag a variable, defaulting to the CUDA 13 image that runs.
- Patch hostNetwork and the LOCALAI_P2P_TOKEN secretKeyRef (Secret
  localai/localai-p2p) onto the Deployment after Helm. Both were set by
  hand, and applying the role would otherwise revert them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

defaults/main.yml

@@ -8,3 +8,16 @@ localai_api_keys:
 localai_watchdog_idle: "false"
 localai_watchdog_idle_timeout: "15m"
 localai_watchdog_busy: "true"
+localai_image_tag: latest-gpu-nvidia-cuda-13
+localai_debug: "false"
+localai_p2p: "false"
+localai_federated: "false"
+# Persistent state lives on the models PVC (mounted at /models).
+localai_data_path: /models/.localai/data
+localai_config_dir: /models/.localai/configuration
+localai_backends_path: /models/.backends
+localai_agent_pool_default_model: ornith-1.5-9b-q8
+localai_agent_pool_embedding_model: granite-embedding-107m-multilingual
+localai_host_network: true
+# Name of a Secret holding LOCALAI_P2P_TOKEN; empty skips the patch.
+localai_p2p_token_secret: localai-p2p