unload_model answers loaded:false while the LLMInferenceService remains #242

Problem

unload_model answered "loaded": false for qwen3-5-4b-hf (backend kserve), but the LLMInferenceService stayed in place and Ready, with its predictor pod running. A delete_node_pool right after it was refused because the pool still served that model:

node pool <pool> still runs 1 busy node(s): node <node> runs model-serving/qwen3-5-4b-hf-kserve-… (1 GPU),
serving 1 model(s): LLMInferenceService model-serving/qwen3-5-4b-hf (Qwen/Qwen3.5-4B) — unload them first

kubectl get llminferenceservice -n model-serving then showed qwen3-5-4b-hf Ready, aged 9m48s: it was the object the preceding load_model had created, so it had never been deleted. On the same pool, an earlier unload_model of the same model removed it as expected, so the failure is intermittent. Observed on an AWS test installation, no model cache configured. The answer carried inventory.reason: "no cache claim is configured; there is no inventory to rescan".

Proposed solution

Have unload_model derive loaded from the serving objects after its delete: the LLMInferenceService gone, or carrying a deletionTimestamp, answered as status: Terminating. A delete that was not issued or that failed must surface as an error, never as loaded: false.

Acceptance criteria

• unload_model answers loaded: false only when the LLMInferenceService is gone or terminating, and otherwise an error naming the object.
• A test covers an unload whose delete does not take effect.